Coding Agents
Long-term project memory for coding agents, backed by Hindsight. One package, several agents: a shared reflect-and-inject core with a thin entry point per agent (Claude Code, Codex CLI, DeepAgents Dcode, opencode, opencode 2, Kilo CLI, Cursor CLI, GitHub Copilot CLI, Grok Build, Qwen Code, Factory Droid, ZCode, Antigravity CLI, Devin CLI, Cline CLI, pi, Prime Agent, DeepSeek Harness). Ingestion is fully automatic — there is no setup command: a repo's git history and conversations flow into its memory bank in the background as you work.
The premise: most of a real fix is derivable from the code, but the last mile often hinges on a project-specific decision that isn't in the code at all — a rounding rule, a retry allowlist, a tie-break policy. Those decisions live in git history and past conversations. This package puts them in front of the agent at the moment it starts working, and keeps a curated set of knowledge pages (architecture, conventions, in-flight initiatives) that future sessions start from.
Install
npx @vectorize-io/hindsight-coding-agents install all # every detected agent, wired natively
npx @vectorize-io/hindsight-coding-agents install claude-code # or just one
npx @vectorize-io/hindsight-coding-agents uninstall all # removes exactly what install added
npx @vectorize-io/hindsight-coding-agents update # refresh the runtime only, no rewiring
npx @vectorize-io/hindsight-coding-agents stats # how often each agent uses Hindsight
install takes an explicit target — all, or one or more harness names. A bare
npx @vectorize-io/hindsight-coding-agents install changes nothing and prints the choice, so wiring every agent on
the machine is never something that happens by accident. Updating is the same install
command again — it re-copies the runtime in place.
Day to day you should not have to: once a day, a session start checks npm and re-stages a newer
runtime in the background (autoUpdate, on by default — set it to false to pin the version you
have). That is the update command above, which refreshes the copy every wired agent already
points at and deliberately touches no host config; re-run install yourself after a release that
adds a new hook, or to wire another agent.
On a terminal it also asks where memory should live — Hindsight Cloud, a server you run, or a
local daemon on this machine (see Where memory lives). Scripted installs pass
--server cloud|self-hosted|daemon instead; it is asked only once, and never again on re-install.
Per agent
Same command, only the harness name changes. Run after installing the package globally.
Claude Code
npx @vectorize-io/hindsight-coding-agents install claude-code
3 hooks in ~/.claude/settings.json, MCP via claude mcp add (user scope), and the companion skill.
Codex CLI
npx @vectorize-io/hindsight-coding-agents install codex
3 hooks in ~/.codex/hooks.json plus [mcp_servers] in config.toml (needs codex_hooks = true).
DeepAgents Dcode
npx @vectorize-io/hindsight-coding-agents install dcode
The installer registers this package as a local marketplace with Dcode, then invokes Dcode's own
plugin install command. The package is a native Agent Plugin: its root plugin.json contributes the shared skill, the
Hooks V2 SessionStart, UserPromptSubmit, and Stop lifecycle, and the namespaced
hindsight_* MCP server. Enable the plugin through Dcode's normal plugin manager; no Dcode config
patcher or compatibility bridge is required.
Dcode namespaces a plugin's MCP tools, so they appear as
plugin__hindsight-coding-…__hindsight_… rather than under their bare names — the agent resolves
them from the tool guide either way. In headless runs (dcode -n) Dcode allows the read-only
Hindsight tools and gates the two that write (hindsight_ingest_document,
hindsight_capture_initiative) behind an approval it has no UI for; use the interactive TUI to
capture an initiative or ingest a document. For the same reason the one-time codebase survey runs
under another installed agent's CLI when there is one, exactly as it does for Cursor, Copilot,
Devin, Grok Build, Cline, Kilo and Prime Agent.
opencode
npx @vectorize-io/hindsight-coding-agents install opencode
A plugin entry in ~/.config/opencode/opencode.json — native tools, no MCP needed.
opencode 2
npx @vectorize-io/hindsight-coding-agents install opencode2
opencode v2 (npm @opencode/cli) exposes an opencode2 binary alias alongside opencode and
rewrote the plugin API, so it is a harness of its own. It writes the same plugin entry to the same
~/.config/opencode/opencode.json — the two CLIs share that file, and v1 rejects the whole config
if it sees v2's plugins key — and each CLI then loads its own entry point from the one registered
path. So installing either harness wires both, and uninstalling either removes the shared entry.
Two differences from v1, both because of the host: the one-time codebase survey runs under another
installed agent's CLI (v2 plugins cannot define the read-only agent the survey needs), and the seed
banner is written to the plugin log instead of a TUI toast (v2 plugins cannot raise one). Recall,
injection, the native hindsight_* tools and session write-back are identical. The companion skill
is identical too, but it is registered in memory through v2's skill API rather than copied into a
skills directory — opencode2 has none of its own, and the two it reads belong to other agents.
Kilo CLI
npx @vectorize-io/hindsight-coding-agents install kilo
A plugin entry in ~/.config/kilo/kilo.json[c].
Cursor CLI
npx @vectorize-io/hindsight-coding-agents install cursor-cli
Hooks in ~/.cursor/hooks.json, ~/.cursor/mcp.json, and the companion skill.
GitHub Copilot CLI
npx @vectorize-io/hindsight-coding-agents install copilot-cli
~/.copilot/hooks/, mcp-config.json, and the companion skill.
Grok Build
npx @vectorize-io/hindsight-coding-agents install grok-build
Native hooks and MCP in ~/.grok/config.toml, plus the companion skill.
Qwen Code
npx @vectorize-io/hindsight-coding-agents install qwen-code
Native hooks in ~/.qwen/settings.json, plus MCP and the companion skill.
Qwen's hook
timeoutis in milliseconds (its own docs: "Timeout in milliseconds, default 60000"), unlike every other supported agent, so the installed values are30000/30000/60000. Recall fires on genuine submissions only —UserPromptSubmitalso fires on tool-result continuations, so interactive sessions recall once per prompt while headless (qwen -p),serve, SDK and ACP sessions seed and retain but do not recall.
Factory Droid
npx @vectorize-io/hindsight-coding-agents install factory-droid
4 hook registrations in ~/.factory/hooks.json (user scope), including a cancellation-safe
Notification write-back, plus a stdio MCP registration under
mcpServers.hindsight in ~/.factory/mcp.json, and the companion skill in ~/.factory/skills.
The installer touches only JSON files - no Droid CLI round-trip - and refuses to overwrite a
user-managed MCP server already named hindsight. Droid's hook protocol matches Claude Code's
(session_id/transcript_path/cwd in, hookSpecificOutput.additionalContext out). Recall and
injection use the same protocol; write-back also handles Droid's cancellation notification because
Droid does not emit Stop after a cancelled turn.
ZCode
npx @vectorize-io/hindsight-coding-agents install zcode
Three hook registrations plus a stdio MCP server under mcp.servers.hindsight, both in ZCode's own
CLI config ~/.zcode/cli/config.json - never your real Claude Code settings, even though ZCode
embeds the Claude Code agent runtime and speaks its hook protocol. The companion skill goes to
~/.zcode/skills. Config hooks ship disabled, so the installer also sets hooks.enabled to
true; uninstall removes the whole block again when nothing else is registered there, and
refuses to touch an MCP server named hindsight that it did not write.
ZCode's hook
timeoutMsis in milliseconds (installed values30000/30000/60000), andhooks.maxOutputBytescaps what a hook may print - anything larger is dropped, injection and all. The installer seeds it at32768only when your config does not already set one.ZCode keeps no durable session transcript:
Stopcarries the reply plus a temp, assistant-only file it deletes as soon as the hook returns, and no user prompt at all. So this is the one agent whose conversation the plugin journals itself - the prompt hook records what you asked, theStophook records the reply - and the write-back then behaves like every other agent's, appending each new turn to the same session document.--import-conversationsis therefore not available for ZCode: there is no past history on disk to backfill from.
Antigravity CLI
npx @vectorize-io/hindsight-coding-agents install agy
Lifecycle hooks, MCP, and the Hindsight · <bank> status line.
Devin CLI
npx @vectorize-io/hindsight-coding-agents install devin-cli
Hooks in ~/.config/devin/config.json plus MCP. Needs Node 22.5+ — see below.
Cline CLI
npx @vectorize-io/hindsight-coding-agents install cline-cli
A native plugin via cline plugin install, plus MCP and the companion skill.
pi
npx @vectorize-io/hindsight-coding-agents install pi
An extension entry in ~/.pi/agent/settings.json, plus the companion skill in
~/.pi/agent/skills — native tools, no MCP needed.
This command is the only supported route, for pi and for Prime Agent below. Installing us as a pi
package (pi install npm:@vectorize-io/hindsight-coding-agents) is deliberately not wired: both
hosts read the same pi key of a package's package.json, and that key can only name one entry —
whichever host it did not name would load the other's bundle and report itself as the wrong agent,
taking that harness's config section and stamping every document it retains with it. So the package
carries no pi key at all, and each host is pointed at its own bundle by the install command above.
Prime Agent
npx @vectorize-io/hindsight-coding-agents install prime-agent
Prime Agent is a fork of pi, so it is wired the same way: an extension entry, here in
~/.prime/agent/settings.json, plus the companion skill in ~/.prime/agent/skills — native tools,
no MCP needed. Installing both is fine and expected:
each host loads its own entry from its own settings file, and like every other pair of agents they
share one bank per repo (the default coding-agent::{gitProject}), so what you tell pi is there
when you open Prime Agent. Separate entries are what keeps each side attributable — its own
harnesses.<name> config section, and its own agent stamped on every document it retains.
DeepSeek Harness
npx @vectorize-io/hindsight-coding-agents install dsh
A Cordis plugin row in $DSH_HOME/cordis.patch.yml (~/.dsh by default), which every dsh profile
composes — native tools, no MCP needed. Two dsh-specific notes: one dsh process serves several
repositories (its Web UI opens each session in whatever directory you pick), so the bank is
resolved per session workspace rather than once per process; and dsh has no plugin-facing notice
channel, so the seed line goes to the plugin log rather than the UI. Everything model-facing —
recalled memory, the knowledge preamble, the hindsight_* tools — is unaffected. If you prefer the
published-package route, dsh plugin --profile web add @vectorize-io/hindsight-coding-agents works
too: the package ships the profile patch layer, so nothing else needs editing. Either route gets the
companion skill — a plugin wired by the host's own plugin manager installs it itself on the first
session, since that route never runs our installer.
Uninstall the same way: npx @vectorize-io/hindsight-coding-agents uninstall claude-code (or uninstall all).
Devin CLI needs Node 22.5 or newer. Its hooks pass only a session id — the conversation itself
lives in ~/.local/share/devin/cli/sessions.db — so reading it depends on Node's built-in
node:sqlite. Installing devin-cli checks for this first and refuses (with the reason) rather
than wiring hooks that could never retain anything. Every other agent works on any supported Node.
install copies what it needs into ~/.hindsight/coding-agents and points each agent's wiring
there, so nothing depends on where you ran it from. Updating is the same command again — it
re-copies the runtime in place, leaving the wiring valid and every new session on the new version.
install merges the native wiring (hooks + MCP registration where the host wants them) into each
agent's own config, preserving everything already there; it is idempotent (re-run after moving the
package) and backs up any pre-existing file it touches as <file>.hindsight-backup. uninstall
removes only our entries. On Claude Code the install also ships a companion skill
(~/.claude/skills/hindsight-coding-agent) that teaches the agent how this memory works — what
"store this in hindsight" should do, the tool surface, per-repo configuration, debugging — so users
can ask the agent itself. Manual wiring per harness, if you prefer:
opencode and opencode 2 install directly — point opencode.json at the package dir:
{ "plugin": ["/path/to/hindsight-coding-agents"] }
One entry, both CLIs: v1 resolves that directory through package.json main, v2 through its
index.js, so each loads its own plugin.
Claude Code and Codex get their full three-hook + MCP wiring from this package's own
installer — npx @vectorize-io/hindsight-coding-agents install claude-code / install codex. This package's bin
entries (hindsight-claude-hook, hindsight-codex-hook,
hindsight-cursor-hook) are the individual injection-only UserPromptSubmit entrypoints for a
minimal, hand-wired setup.
Adding an agent: hook-based → write a HookSpec entry point (see src/cursor-hook.ts) and register
a hookAdapter in src/harness/registry.ts; persistent-plugin → implement HarnessAdapter
(src/core/types.ts) fully (see src/harness/opencode.ts), or bind the host's own plugin API to
RuntimeCore directly when it is not an opencode fork (see src/cline.ts, src/dsh.ts).
Migrating from the per-agent plugins
The older per-agent integrations (hindsight-claude-code, hindsight-cursor-cli, hindsight-codex, …) are superseded by this package. Two things move; nothing else does.
Your server moves automatically. If ~/.hindsight/claude-code.json or ~/.hindsight/codex.json
exists, install adopts its endpoint — hindsightApiUrl → apiUrl, hindsightApiToken →
apiToken, and an empty URL means the local daemon, as it did there. The agent you are installing
is checked first, so wiring Codex takes Codex's server even if an old claude-code.json is still
lying around. You already chose where your memory lives; defaulting to Cloud instead would quietly
send your prompts somewhere else. Pass --server to override. (Those two are the only old plugins
that shipped a user config — Cursor CLI, Copilot CLI, opencode and Cline have no endpoint to carry.)
Your conversations are re-imported from local disk, as new documents:
cd /path/to/your/repo
npx @vectorize-io/hindsight-coding-agents install claude-code --import-conversations # or: install codex --import-conversations
This re-extracts the transcripts the agent already wrote, so it costs tokens roughly in proportion to the history imported, and it is safe to re-run (ingestion dedups by document id).
Local transcripts are the source rather than the old bank, because the old bank cannot be split by
repo. Its default was a single static bank — dynamicBankId defaults to false, so everything
landed in one bank named claude_code — and its documents record only retained_at,
message_count and session_id, nothing identifying the project. Working out which documents
belong to which repo means joining session_id back to the cwd in the local transcript, so the
transcripts are needed either way; going through them directly is simply the shorter path.
How sessions are matched. A conversation is imported only when the session itself records the
directory it ran in — never inferred from a file or folder name. Claude Code writes that directory
on its entries, Codex in its session_meta header and DeepSeek Harness in its session-log header,
and pi and Prime Agent in their session header, so all five can be attributed exactly, including
sessions started in a subdirectory of the repo. Guessing was tempting (Claude names its history
folders after the project path) but unsafe: / and . both encode to -, so repo-sub is either
the subdirectory repo/sub or an unrelated sibling repo — and a wrong guess files someone else's
conversation into your bank. Sessions that record nothing are skipped and the count is reported.
DeepSeek Harness logs are Zstandard-framed JSONL under $DSH_HOME/sessions, which needs Node 22.15+
to read; an older Node skips the import with that reason rather than silently importing nothing.
Dcode's transcripts record no directory at all — the working directory lives only in its LangGraph
checkpoint database — so the repo comes from dcode threads list --json, a declared, versioned
command contract rather than that internal schema; with the dcode CLI unavailable the import is
skipped with that reason. The other harnesses (opencode, opencode 2, Kilo, Cursor, Cline, Copilot, Devin) keep
history in internal SQLite databases with unversioned schemas and are skipped with a reason.
Nothing else is translated. The old plugin's behavioural settings — 12 recall*, 7 retain*,
bankMission/retainMission, dynamicBankGranularity — describe a pipeline this package replaced,
and reinterpreting them would be guesswork. Bank naming changes too: this package uses one bank per
repo (coding-agent::{gitProject}) shared by every agent. To keep the old naming instead:
{ "bankIdTemplate": "{harness}::{gitProject}" } // reproduces the old per-agent naming
Where memory lives
Three modes, chosen once when you install (install asks on a terminal; pass --server to script it):
| mode | what runs | needs |
|---|---|---|
cloud | Hindsight Cloud (default) | an API token |
self-hosted | a Hindsight server you already run | its URL |
daemon | a local hindsight-embed on this machine | uv on PATH + an LLM key for extraction |
npx @vectorize-io/hindsight-coding-agents install claude-code --server daemon
npx @vectorize-io/hindsight-coding-agents install claude-code --server self-hosted --api-url http://localhost:8888
npx @vectorize-io/hindsight-coding-agents install claude-code --server cloud --api-token <token>
Re-running install never re-asks: a config that already names a server is left alone.
Local daemon mode
Nothing to sign up for and nothing to host — memory runs on your machine. The plugin starts
hindsight-embed on demand at 127.0.0.1:9077 and points every agent at it.
- A server already on the port is adopted, never restarted — so one daemon serves every agent
and every repo, and your own
hindsight-embedis reused if you already run one. - Cold starts happen in the background. The first start downloads the daemon and loads models, which takes longer than any hook is allowed to run, so it is launched detached at session start. A session that begins before it is ready simply has no memory for a turn or two — a daemon that isn't up is treated as an unreachable server, exactly like a Cloud or self-hosted outage, with the same error handling and the same diagnostics. Nothing downstream of the URL knows which mode it is.
- It keeps running until you stop it. There is deliberately no stop-on-exit: one daemon is shared, so ending one session must not cut memory out from under another agent still working.
- macOS additionally needs a current Rust toolchain.
litellm(a transitive dependency of the API) publishes wheels only for Linux and Windows, so a Mac compiles it from source through maturin and its crates pin a recentrustc. Install from rustup.rs and keep it updated — an out-of-date toolchain fails as surely as a missing one. Linux and Windows install from wheels and need none of this. - Fact extraction runs locally, so it needs an LLM.
HINDSIGHT_API_LLM_PROVIDERwins if set; otherwise the first ofOPENAI_API_KEY,ANTHROPIC_API_KEY,GEMINI_API_KEY,GROQ_API_KEY; otherwise the Claude Code CLI, which needs no key.installtells you which it found.
Daemon settings keep the names the old per-agent Claude Code plugin used, so an existing environment carries over unchanged:
| field | env | default | meaning |
|---|---|---|---|
serverMode | HINDSIGHT_SERVER_MODE | cloud | cloud | self-hosted | daemon |
apiPort | HINDSIGHT_API_PORT | 9077 | port the local daemon listens on |
daemonIdleTimeout | HINDSIGHT_DAEMON_IDLE_TIMEOUT | — | deprecated, ignored: the daemon no longer exits on its own |
daemonProfile | HINDSIGHT_DAEMON_PROFILE | coding-agent | which local database it uses |
embedVersion | HINDSIGHT_EMBED_VERSION | latest | which hindsight-embed release to run |
embedPackagePath | HINDSIGHT_EMBED_PACKAGE_PATH | — | run a local checkout instead (development) |
Any HINDSIGHT_API_* variable you export is forwarded to the daemon, so server-side settings need
no equivalent here.
Configuration
Configuration is one JSON file: ~/.hindsight/coding-agent.json. Layering, later wins per field:
- built-in defaults
- environment variables —
HINDSIGHT_API_URL,HINDSIGHT_API_TOKEN, and one per scalar setting (HINDSIGHT_<FIELD_IN_CAPS>), for containers and CI that inject config rather than write a file - the file's top level
- its
harnesses.<name>section — per-agent override - its
banks.<resolvedBankId>section — per-repo override, applied after the bank is resolved (see Per-repo opt-in/out)
Environment variables are a fallback: the file wins wherever it sets a value, so adding env to
an existing setup changes nothing. The two list-valued settings, retainTags and optInPaths, take
a comma-separated value (HINDSIGHT_RETAIN_TAGS="project:{gitProject},env:work"); entries are
trimmed and blanks dropped.
The map-valued settings (mapPathToBank, harnesses, banks, retainMetadata) are file-only —
per-key branching doesn't survive flattening into one variable. maxParallelRetains is available
as HINDSIGHT_MAX_PARALLEL_RETAINS for containers and CI.
HINDSIGHT_CONFIG moves the file itself — point it at another path for a container or a test
harness where $HOME is not the right anchor. It is still exactly one file; only its location
changes. (The other variables that are not settings are HINDSIGHT_LOG_FILE, HINDSIGHT_DIAG_FILE,
HINDSIGHT_USAGE_FILE and HINDSIGHT_LOG_LEVEL — see Diagnostics & logging.)
When a change takes effect
Config is read when a process starts — the file is not watched — so when an edit applies depends on what reads it:
| host | reads the file | an edit applies |
|---|---|---|
| hook harnesses (Claude Code, Codex CLI, Cursor CLI, GitHub Copilot CLI, Grok Build, Antigravity CLI, Devin) | once per hook invocation — each hook is its own short-lived process | on your next prompt |
| persistent plugins (opencode, opencode 2, Kilo CLI, Cline CLI, pi, Prime Agent, DeepSeek Harness) | once per workspace, when the host loads the plugin | after restarting the agent |
the MCP server behind the hindsight_* tools | once at startup | in your next session |
apiToken is the exception. Every host re-reads it when the server rejects a request, so enabling
authentication or rotating the key is picked up on the next call with nothing to restart —
otherwise a rotation would leave a long-running agent failing every memory call until it was
restarted. Everything else follows the table: apiUrl, disabled, bank routing, gitIngest, and
the survey and knowledge-page settings.
hindsight_diagnose reports both sides of that gap — what the file says now, and what the running
client is actually using.
Opt-in only
By default every project gets memory — that is what makes the plugin zero-setup. If you would rather nothing be remembered until you say so, turn memory off everywhere and name the projects that may use it:
{
"optInOnly": true,
"optInPaths": ["~/work/client-x", "~/oss"],
}
Anything outside those paths is inert: no bank is created, nothing is retained, no seed runs,
and the agent behaves exactly as it would without the plugin. Approving costs nothing else —
optInPaths says which projects, not which bank, so an approved repo keeps its usual
coding-agent::{gitProject} name. Paths are prefixes, so approving ~/work approves every repo
under it while each still gets its own bank.
A mapPathToBank entry counts as opted in too, since routing a path to a named bank already
declares that project. A bare bankId does not: it names a bank rather than a project, so it
cannot say which work may be remembered, and a privacy switch has to fail closed.
There is no per-repo opt-in file, for the same reason there is no repo-carried config at all: a cloned repository must not be able to turn memory on.
There is deliberately no repo-carried config file — per-repo bank routing is mapPathToBank,
per-agent differences are harnesses.<name>.
Each entry point knows which harness it is (the opencode plugin is loaded by opencode, the codex hook by Codex...), so one shared config serves several agents side by side:
{
"apiUrl": "https://api.hindsight.vectorize.io",
"harnesses": {
"opencode": { "reflectTimeoutMs": 60000 },
"claude-code": { "disabled": true }, // e.g. memory off for Claude only
},
}
Reference
| field | default | meaning |
|---|---|---|
apiUrl | https://api.hindsight.vectorize.io | Hindsight API base URL (set to http://localhost:8888 for a local server) |
apiToken | — | bearer token (Hindsight Cloud). Picked up without restarting the agent: a long-lived host re-reads it after a rejected request, so enabling auth or rotating the key mid-session recovers on the next call |
bankId | — | explicit static bank; unset ⇒ per-repo dynamic resolution (below) |
dynamicBankId | dynamic iff no bankId | force dynamic (true) or static (false) resolution |
bankIdTemplate | "coding-agent::{gitProject}" | dynamic bank id format; the default makes every agent share one bank per repo |
mapPathToBank | — | absolute path → bank; longest prefix wins; linked worktrees inherit their main checkout's mapping; overrides everything |
optInOnly | false | run memory ONLY in opted-in projects — everything else is inert, with no bank created; see Opt-in only |
optInPaths | — | directories opted in, matched as prefixes with ~ expanded; each repo beneath and its linked worktrees are approved while keeping their own dynamic bank |
resolveWorktrees | true | linked worktrees inherit the main checkout's bank identity, path approval, and mapping |
retainTags | — | extra tags on every document written by the integration, e.g. ["project:{gitProject}"] — see Recording where a memory came from below |
retainMetadata | — | extra metadata on every document written by the integration, e.g. {"repo": "{gitProject}"} |
manageBankConfig | true | let the plugin shape the bank's own configuration — the retain strategies it writes under, the knowledge entity-label group, and, on a bank that has none, the missions. Writing is strictly additive: it adds what the bank does not define and never overwrites what is there, so your control-plane edits survive. Set false to keep it out of the bank config entirely — see A bank you shape yourself below |
observationScopes | "shared" | how consolidation groups observations: "shared" (default) = ONE global scope per bank, so every agent on a repo builds one set of beliefs; also "combined" (the server default), "per_tag", "all_combinations", [["t"]]; "per_source" adds a scope per source: kind alongside the global one, so commit knowledge and conversation knowledge consolidate apart |
disabled | false | hard off-switch (inert plugin/hook — a no-memory baseline) |
reflectTimeoutMs | 20000 | automatic session-reflect timeout; on hook harnesses the installer registers a 30s prompt-hook timeout, so going above ~20s also means raising that hook's timeout in the host's config, or the host kills the hook mid-reflect; on timeout or a 5xx the hook falls back to knowledge-page search, then to a raw recall of observations (recorded) |
reflectToolTimeoutMs | 330000 | timeout for the agent-invoked hindsight_reflect tool — a call the agent waits on, whose high-budget synthesis on a populated bank runs for minutes. Defaults above the server's own reflect wall timeout (HINDSIGHT_API_REFLECT_WALL_TIMEOUT, 300s) so the server decides when to give up. Unset, it inherits an explicitly raised reflectTimeoutMs, but a short one never lowers it |
reflectBudget | "high" | reflect budget for the hindsight_reflect tool: "low", "mid" or "high". Drop it on a large bank where high-budget synthesis exceeds the server's wall timeout. The automatic session-start reflect always uses "low" to fit its hook window and is unaffected |
autoInject | "reflect" | what to inject once, on the session's first prompt: "reflect" = a low-budget reflect synthesis (on timeout/5xx it falls back to page search, then recall); "pages" = the knowledge pages matching the prompt by search; "recall" = the bank's memories recalled for the prompt (what it asks for is recallOptions, observations by default); "none" = nothing — the agent searches knowledge pages first and reflects only when they are too shallow. "pages" and "recall" are retrieval only (no LLM), so they stay well inside the hook window |
autoReflect | true | deprecated — use autoInject. Still honoured (false = autoInject: "none"), ignored when autoInject is set, and logs a deprecation warning |
pageSearchLimit | 3 | knowledge pages ONE search returns. Applies to every search of the bank — the autoInject: "pages" injection, the reflect fallback, and the agent's own hindsight_search_knowledge_pages tool — because the limit lives on the client they all share |
recallOptions | see description | overrides merged key-by-key into the body of every recall — the autoInject: "recall" source and the reflect fallback. Keys are the API's own recall parameters, passed straight through, so anything recall accepts is settable without a new option here; query is the one field it cannot replace. The default body is {"types": ["observation"], "budget": "low", "max_tokens": 2000, "include": {"entities": null}}, and what you set is merged over it one key at a time — {"max_tokens": 4000} changes the budget and leaves the rest alone. Observations are the consolidated layer, so they answer best per token, but a bank with consolidation disabled never grows any and the default recall comes back empty on it: set {"types": ["world", "experience"]} there, or {"types": null} for every type. File-only, like retainMetadata — an object does not flatten into an env var |
pageRefreshEveryTurns | 10 | refetch the knowledge pages and re-inject the page roster + tool guide every N user turns |
pageTriggerType | "cron" | when NEW knowledge pages refresh, i.e. what keeping them current costs — "cron" (default) on pageTriggerCron only and only when actually stale, "auto-refresh" after every consolidation that produced new material, "manual" never on their own. Auto-refresh is the most current and by far the most expensive: one synthesis per page per consolidation. Maps to the page's trigger.refresh_cron, or trigger.refresh_after_consolidation in the Hindsight API (true for auto-refresh, false for manual) |
pageTriggerCron | "H * * * *" | schedule for pageTriggerType: "cron" — UTC, standard 5-field cron, e.g. "0 3 * * *". The default is hourly, each page on its own hashed minute. Sets the page's trigger.refresh_cron, which the API treats as mutually exclusive with refresh_after_consolidation; a scheduled refresh is skipped when nothing changed. Write a field as H to give each page its own value there — see Spreading refreshes with H below |
autoSeed | true | SessionStart: auto-seed a cold repo's bank from git history |
seedLimit | 300 | auto-seed: most-recent-N-commits cap |
codebaseSurvey | true | SessionStart: headless survey of a cold repo's structure, run under the current harness's own CLI (claude/codex/antigravity/opencode), falling back to any available agent |
surveyModel | haiku | model for the survey — Claude recipe only (claude -p --model); other agents use their configured default |
surveyBudgetUsd | 2 | survey spend cap — Claude recipe only (claude -p --max-budget-usd); other agents rely on their read-only sandbox |
surveyRefreshCommits | 20 | re-run the survey at SessionStart once this many commits have accrued since the last one, so the structural pages track an architecture that keeps moving (0 = survey a cold repo only, never again) |
retainSessions | true | session write-back, honored by every harness: hook harnesses write the transcript on Stop, Factory Droid also writes on its cancellation notification, and plugin harnesses (opencode, opencode 2, Kilo) upsert it every turn plus an idle flush that captures the reply the per-turn pass can't see. Set false - globally, per harness, or per bank - to stop writing transcripts (the background history import stops with it) while recall, git ingest and the memory tools keep working |
maxParallelRetains | 10 | cap on concurrent retain-related requests: drain()'s per-op polls plus deepen's chat/git retain pools. The API rate-limits bursts, not single requests — if you see 429s, lower this rather than raising it |
logLevel | "info" | plugin-log verbosity ("debug" | "info" | "warn" | "error"); HINDSIGHT_LOG_LEVEL env overrides |
autoUpdate | true | keep the installed runtime current by itself: once a day a session start asks npm for the published version and, when it is newer, re-stages ~/.hindsight/coding-agents in the background. It rewires no host config, so a release adding a new hook entry point still needs a manual install. Set false to pin the installed version; disabled stops it too, since an inert plugin should stay inert. Only ever replaces a runtime installed the documented way, via npx — a copy installed with npm i -g, vendored as a project dependency, or built from a checkout is left to whoever manages it (update those the way you installed them), and it needs npx and npm on PATH |
gitIngest | "message" | git depth for seeding AND staying current (same engine): "message" = commit messages only (one doc, re-upserted when HEAD moves); "full" = messages + per-commit full diffs (progressive, newest first); "none" = git off |
harnesses.<name> | — | per-harness override of any field above |
harness | opencode | deepen engine only: which session format --conversations is read as |
By default a page refreshes hourly, staggered: pageTriggerCron is "H * * * *", so every
page gets its own minute of the hour (see below) and a tick with nothing new to fold in is skipped
server-side. That keeps pages within an hour of the repo without paying auto-refresh's price — one
LLM synthesis per page per consolidation, on a repo that consolidates all day. Set
pageTriggerType: "auto-refresh" to go back to refreshing on every consolidation.
pageTriggerType/pageTriggerCron decide only when a page refreshes. How it refreshes
belongs to the server: Hindsight creates a knowledge page with a delta refresh (each pass edits the
page instead of rebuilding it) that doesn't reflect over sibling pages, and these settings merge
over those defaults rather than replacing them.
Spreading refreshes with H
One pageTriggerCron is shared by every page in every repo you point this plugin at. So a literal
"0 3 * * *" does not schedule a refresh at 03:00 — it schedules all of them at 03:00, five
pages per bank, on the same worker pool that serves retain. A session ingesting at 03:0x queues
behind the pile, and moving the hour just moves the pile.
Write a field as H and it is replaced, per page, by a value hashed from the bank id and the page
name. Each page gets its own slot, the same slot on every run:
pageTriggerCron | what each page gets |
|---|---|
"H H * * *" | once a day, at its own minute and hour |
"H * * * *" | once an hour, at its own minute |
"H 3 * * *" | daily at 03:MM — spread inside the hour you chose |
"H H(0-5) * * *" | daily, spread across 00:00–05:59 only |
"0 3 * * *" | no H, no hashing — exactly what it says, all at once |
H is Jenkins' syntax for the same
problem. It never reaches the API: the plugin resolves it to an ordinary cron expression
("41 17 * * *") when it creates the page, so the schedule you see in the control plane is a plain
one you can edit. Hashing spreads pages out, it does not partition them — two pages can still land
on the same minute, just not all of them.
These settings apply to the pages a repo already has, too. Every session compares each page
this plugin created — the seeded taxonomy and every captured initiative — against the config and
re-syncs the ones that differ, so a bank seeded before this default changed moves onto the hourly
schedule by itself, and a page you retriggered by hand in the control plane is put back on the
configured policy the next time an agent runs. The config
file is the source of truth for these pages: to give one a different schedule, change
pageTriggerType/pageTriggerCron (per bank, if it is only that repo) rather than editing the
page. Only the fields this plugin states are touched — a page's mode, its excluded siblings and
its minimum refresh interval are left exactly as they are.
A bank you shape yourself — manageBankConfig
Pointed at a bank, this plugin gives it the shape its ingestion needs: retain strategies for the
kinds of document it writes (git, gitlog, conversation, document, survey), a knowledge
entity-label group that routes facts to the knowledge pages, and — on a bank that has no missions of
its own — the coding missions.
It only ever adds what is missing. A strategy you defined, an edit you made to one of the
plugin's, a reworded label group, a mission you rewrote in the control plane: each is left exactly
as it is, on every session, forever. What the bank already says wins. The cost of that promise is
that a plugin release which rewords an existing strategy or label does not reach a bank that
already has it. To take the current default back, clear that override on the bank (delete the
strategy, or the whole retain_strategies entry, in the control plane): the next session finds the
bank silent there and seeds it again.
Set manageBankConfig: false to keep the plugin out of the bank's configuration altogether — the
right setting for a bank you share with non-coding work, or one you configure yourself. That bank
should then define the five strategies above itself. Note that the miss is silent: the server
does not reject a retain naming a strategy the bank lacks, it logs a warning and extracts with the
bank's own configuration — so a commit diff, a session transcript and a survey marker would all get
the same generic treatment instead of the extraction each needs. Knowledge pages are seeded either
way; pageTriggerType governs what they cost.
Like every field here it can be set per bank, which is usually where it belongs:
{
"bankId": "my-global-bank",
"banks": { "my-global-bank": { "manageBankConfig": false } }
}
Per-repo opt-in/out — banks.<bankId>
Per-repo control lives in the SAME file, keyed by the resolved bank id (shown in the session banner) and applied AFTER bank resolution — so it works regardless of where the repo lives, and survives directory moves:
{
"banks": {
"coding-agent::secret-client": { "disabled": true }, // blacklist: no memory at all
"coding-agent::old-name": { "bank": "team::shared" }, // rename / converge banks
"coding-agent::big-mono": { "gitIngest": "full", "retainSessions": false },
},
}
Any behavioral field can be overridden per bank, and bank renames the destination (single
hop: the section is selected by the resolved id, the target is literal — several ids may converge
on one shared bank, and the target's own section is not consulted). Other bank-resolution fields
are ignored inside a bank section.
Recipe: two repos, one shared bank
Two ways, by what the natural key is:
By resolved id — you know the repo names; works wherever the repos live (and keeps working if they move). Both ids converge on one literal target:
{
"banks": {
"coding-agent::backend": { "bank": "team::product" },
"coding-agent::frontend": { "bank": "team::product" },
},
}
By path prefix — the repos live under one directory; a single mapPathToBank entry covers
every repo (present and future) beneath it:
{
"mapPathToBank": { "/Users/me/work/client-x": "client-x-memory" },
}
Rule of thumb: converge by id for a hand-picked set of repos; map by path when a folder is
the boundary ("everything I clone under work/client-x shares memory").
Bank resolution
Coding memory is per repository. Resolution order for the working directory:
mapPathToBank— longest matching absolute-path prefix (mapping a repo root covers every subdirectory; deeper mappings win; overrides even an explicitbankId).- Static —
bankIdset (ordynamicBankId: false). - Dynamic —
bankIdTemplatewith placeholders:{gitProject}— worktree-aware repo name:git rev-parse --git-common-dirresolves every linked worktree to the main worktree's basename, so all worktrees of a repo share one bank (bare repos use the bare dir name). Outside a repo there is nothing for git to resolve, so it falls back to the basename of the directory the session started in — an agent thatcds into a subdirectory keeps writing to one bank, and a subdirectory gets its own bank only when you deliberately start a session there{project}— plain working-directory basename{harness}— the entry point asking (opencode,claude-code,codex,antigravity-cli,cursor-cli,copilot-cli){channel}/{user}—$HINDSIGHT_CHANNEL_ID/$HINDSIGHT_USER_ID
The default "coding-agent::{gitProject}" is harness-neutral, so opencode, Claude Code, and Codex
all share one memory per repo — use "{harness}-{gitProject}" to split per agent instead.
Recording where a memory came from
With a bank per repo, the bank is the answer to "where did this come from". On a deliberately
shared bank — one bank holding cross-project knowledge so facts recall everywhere — it isn't:
every memory looks alike. retainTags and retainMetadata stamp that provenance onto conversations,
git history and diffs, survey lifecycle documents, initiative markers, and documents saved through
hindsight_ingest_document:
{
"bankId": "shared", // one bank for everything
"retainTags": ["project:{gitProject}", "env:work"],
"retainMetadata": { "repo": "{gitProject}" },
}
Recalls can then filter by project:<repo>, and every document shows which repository it came out
of. Both accept the same placeholders as bankIdTemplate — {gitProject}, {project},
{harness}, {channel}, {user} — plus {bankId}, {sessionId} and {timestamp}.
{gitProject} is worktree-aware here too, so every linked worktree of a repo stamps one name.
{sessionId} resolves to unknown for documents that do not originate from an agent session.
The plugin's own source: and harness: tags are reserved: entries in those namespaces are ignored
with a warning, so a document's agent attribution always reflects the agent that actually wrote it.
One set of beliefs per repo
Every document this integration writes carries provenance tags — source:chat, harness:<id>,
knowledge:<kind>, plus anything from retainTags. Those tags say who wrote a memory; they are
what filters recall and draws each document's agent logo, and they stay on the facts.
They are not, however, a good boundary for
observations. Consolidation's own
default (combined) builds one observation set per distinct tag set, so the same repository
worked on by two agents would grow two parallel sets of beliefs — one per harness — that never
merge, each blind to the other, at double the consolidation cost. Which agent happened to be typing
does not change whether a convention or a decision is true.
So the integration retains with observationScopes: "shared": one global, untagged observation
scope per bank, which is what a bank already is — one project's memory. Set the field to change it:
{
"observationScopes": "combined", // one observation set per distinct tag set (server default)
"banks": {
"coding-agent::mono": { "observationScopes": "per_tag" }, // per-repo, like any behavioral field
},
}
Splitting code from conversation — per_source
shared puts every document a repo produces into one belief set. "per_source" keeps that set and
adds one per origin, so "what the commits say" and "what was decided in conversation" can be asked
apart:
{ "observationScopes": "per_source" }
Each document consolidates into the global scope plus one named for each source: tag it
carries — [[], ["source:chat"]] for a session transcript, [[], ["source:git"]] for a commit
diff. Read an axis back with tags: ["source:git"], tags_match: "exact", and the merged view with
tags: [], tags_match: "exact".
A document carrying two source: tags gets a scope for each, and that is deliberate rather than
duplication. The commit-message seed is tagged source:git and source:git-log, so
source:git-log is fed only by the seed — what the commit messages say — while source:git also
collects every per-commit diff under gitIngest: "full". Two questions, two answers, each
deduplicated within itself by consolidation. A fact belonging to more than one axis is the point.
This cannot be expressed as a scope list. The server treats an explicit list[list[str]] as
unconditional — it is not filtered against the memory's own tags — so a configured
[[], ["source:git"], ["source:chat"]] writes every document into all three, and the source:git
scope fills with beliefs built from chat transcripts. Only a per-document decision separates them.
It costs one extra consolidation pass per document, and it reads only source:, so a volatile
provenance tag never becomes a scope. The global scope is still written first and unchanged, so the
untagged observations knowledge pages read are unaffected.
"per_tag" and "all_combinations" split further still, and an explicit [["project:demo"], …]
declares the scopes literally. HINDSIGHT_OBSERVATION_SCOPES sets the scalar modes; a scope list is
file-only. Changing this does not rewrite observations already consolidated under the old scoping —
they stay where they were built, and new work accrues under the new setting.
Diagnostics & logging
All logs live in ~/.hindsight/coding-agents-logs/ (owner-only). Each file rotates to <file>.1
at 10 MB.
Leveled plugin log (humans debugging): ~/.hindsight/coding-agents-logs/plugin.log (override
HINDSIGHT_LOG_FILE) — timestamped LEVEL [scope] message lines from every component, including
the ingestion engine. Level defaults to info; set "logLevel": "debug" in config or
HINDSIGHT_LOG_LEVEL=debug for ad-hoc debugging (at debug, every diag event below is mirrored
here too, so one file tells the whole story).
Structured diag events (machines/harnesses): every reflect and page-fetch outcome is appended
as a JSON line to ~/.hindsight/coding-agents-logs/diag.jsonl (override with
HINDSIGHT_DIAG_FILE):
{
"ts": "2026-07-27T07:05:52Z",
"harness": "claude-code",
"event": "reflect_ok",
"ms": 14210,
"chars": 792,
"query": "..."
}
reflect_failed / pages_failed record the error; if you're comparing memory-on vs memory-off,
check this file — a run whose reflects failed is a no-memory run. When the failure was a timeout or
a 5xx, the hook falls back to knowledge-page search and, if no page matches, to a raw recall of the
bank's memories: reflect_fallback_pages / reflect_fallback_observations record what each
step returned (*_failed when it errored). Seed starts are logged as
seed_started.
When autoInject names a retrieval source instead of reflect, that source records its own
outcome with the number of items it returned: inject_pages for "pages", inject_recall for
"recall" (inject_pages_failed / inject_recall_failed when the call errored). No reflect
event is written on those turns — nothing reflected.
Tool usage (is the agent using Hindsight?): one JSON line per finished user turn in
~/.hindsight/coding-agents-logs/usage.jsonl (override HINDSIGHT_USAGE_FILE) — the hindsight_*
tools the agent called during that turn, and whether its reply credited Hindsight memory ("From
Hindsight memory"). The credit rate counts only turns that called a retrieval tool (search, list,
read, reflect); saving a document is not expected to be credited. Recorded when the session is written back, so a scope with
retainSessions: false records none. It never leaves your machine. Summarize it per agent with:
npx @vectorize-io/hindsight-coding-agents stats
Is the memory ready yet?
hindsight_sync_status — the agent-facing tool, dist/status.js for scripts — answers exactly
that: "synced": true means the seeded memory is queryable. It also reports gitlog freshness, how
far per-commit deepening has got, the codebase survey's state (surveyBaseline is the HEAD the last
survey started from, surveyDocs counts the findings documents that have landed, 0–4 — a baseline
with no findings retries automatically), and the extraction operations still in flight.
Resetting a repo's memory
Delete its bank on the server. The bank is the only state this integration keeps, so the next session in that repo is a true first open — seed and survey run again from scratch. There are no client-side files to clean up.
Marker documents you may notice
Two document ids exist for the machinery's own bookkeeping. Both are safe to ignore and safe to delete:
survey-baseline:<sha>— reads "🛰️ researching…" while a codebase survey runs and flips to "✅ completed" once its findings land. It is retained under thesurveystrategy, whose marker rule extracts nothing from a status marker, and it drives the re-survey cadence (surveyRefreshCommits) andsurveyBaselinein sync status.gitlog:<repo>— the aggregated commit-message seed document, re-upserted rather than duplicated when the seed runs again.
When memory seems to be missing
Failures never break the agent: a reflect, page fetch or retain that fails degrades to an ordinary
memoryless turn and is recorded in the logs. "No memory" is therefore a log question — check the
diag file for whether session_start and deepen_started ever fired for that bank. A session that
was already running when the plugin was installed has no SessionStart behind it; its first prompt
after the install self-heals.