@vectorize-io/hindsight-coding-agents 0.1.0 → 0.1.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +93 -28
- package/dist/antigravity-hook.js +4 -2
- package/dist/antigravity-statusline.js +3 -1
- package/dist/antigravity-stop-hook.js +4 -2
- package/dist/claude-hook.js +4 -2
- package/dist/claude-sessionstart-hook.js +5 -3
- package/dist/claude-stop-hook.js +4 -2
- package/dist/cline.js +5 -3
- package/dist/codex-hook.js +4 -2
- package/dist/codex-sessionstart-hook.js +5 -3
- package/dist/codex-stop-hook.js +4 -2
- package/dist/copilot-hook.js +4 -2
- package/dist/copilot-sessionstart-hook.js +5 -3
- package/dist/copilot-stop-hook.js +4 -2
- package/dist/cursor-hook.js +4 -2
- package/dist/cursor-sessionstart-hook.js +5 -3
- package/dist/cursor-stop-hook.js +4 -2
- package/dist/daemon-start.js +3 -1
- package/dist/deepen.js +4 -2
- package/dist/devin-hook.js +4 -2
- package/dist/devin-sessionstart-hook.js +5 -3
- package/dist/devin-stop-hook.js +4 -2
- package/dist/grok-hook.js +4 -2
- package/dist/grok-sessionstart-hook.js +5 -3
- package/dist/grok-stop-hook.js +4 -2
- package/dist/hindsight-seed.js +3 -1
- package/dist/index.js +5 -3
- package/dist/kilo.js +5 -3
- package/dist/mcp-server.js +4 -2
- package/dist/status.js +4 -2
- package/package.json +1 -1
- package/skill/SKILL.md +4 -4
package/README.md
CHANGED
|
@@ -15,14 +15,13 @@ in front of the agent at the moment it starts working, and keeps a curated set o
|
|
|
15
15
|
## Install
|
|
16
16
|
|
|
17
17
|
```bash
|
|
18
|
-
|
|
19
|
-
hindsight-coding-agents install
|
|
20
|
-
hindsight-coding-agents
|
|
21
|
-
hindsight-coding-agents uninstall all # removes exactly what install added
|
|
18
|
+
npx @vectorize-io/hindsight-coding-agents install all # every detected agent, wired natively
|
|
19
|
+
npx @vectorize-io/hindsight-coding-agents install claude-code # or just one
|
|
20
|
+
npx @vectorize-io/hindsight-coding-agents uninstall all # removes exactly what install added
|
|
22
21
|
```
|
|
23
22
|
|
|
24
23
|
`install` takes an explicit target — `all`, or one or more harness names. A bare
|
|
25
|
-
`hindsight-coding-agents install` changes nothing and prints the choice, so wiring every agent on
|
|
24
|
+
`npx @vectorize-io/hindsight-coding-agents install` changes nothing and prints the choice, so wiring every agent on
|
|
26
25
|
the machine is never something that happens by accident.
|
|
27
26
|
|
|
28
27
|
On a terminal it also asks **where memory should live** — Hindsight Cloud, a server you run, or a
|
|
@@ -33,30 +32,96 @@ local daemon on this machine (see [Where memory lives](#where-memory-lives)). Sc
|
|
|
33
32
|
|
|
34
33
|
Same command, only the harness name changes. Run after installing the package globally.
|
|
35
34
|
|
|
36
|
-
|
|
37
|
-
|
|
38
|
-
|
|
39
|
-
|
|
40
|
-
|
|
41
|
-
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
|
|
35
|
+
#### <img src="https://hindsight.vectorize.io/img/harness/claude-code.png" alt="" width="20" height="20" /> Claude Code
|
|
36
|
+
|
|
37
|
+
```bash
|
|
38
|
+
npx @vectorize-io/hindsight-coding-agents install claude-code
|
|
39
|
+
```
|
|
40
|
+
|
|
41
|
+
3 hooks in `~/.claude/settings.json`, MCP via `claude mcp add` (user scope), and the companion skill.
|
|
42
|
+
|
|
43
|
+
#### <img src="https://hindsight.vectorize.io/img/harness/codex.svg" alt="" width="20" height="20" /> Codex CLI
|
|
44
|
+
|
|
45
|
+
```bash
|
|
46
|
+
npx @vectorize-io/hindsight-coding-agents install codex
|
|
47
|
+
```
|
|
48
|
+
|
|
49
|
+
3 hooks in `~/.codex/hooks.json` plus `[mcp_servers]` in `config.toml` (needs `codex_hooks = true`).
|
|
50
|
+
|
|
51
|
+
#### <img src="https://hindsight.vectorize.io/img/harness/opencode.png" alt="" width="20" height="20" /> opencode
|
|
52
|
+
|
|
53
|
+
```bash
|
|
54
|
+
npx @vectorize-io/hindsight-coding-agents install opencode
|
|
55
|
+
```
|
|
56
|
+
|
|
57
|
+
A plugin entry in `~/.config/opencode/opencode.json` — native tools, no MCP needed.
|
|
58
|
+
|
|
59
|
+
#### <img src="https://hindsight.vectorize.io/img/harness/kilo.svg" alt="" width="20" height="20" /> Kilo CLI
|
|
60
|
+
|
|
61
|
+
```bash
|
|
62
|
+
npx @vectorize-io/hindsight-coding-agents install kilo
|
|
63
|
+
```
|
|
64
|
+
|
|
65
|
+
A plugin entry in `~/.config/kilo/kilo.json[c]`.
|
|
66
|
+
|
|
67
|
+
#### <img src="https://hindsight.vectorize.io/img/harness/cursor-cli.svg" alt="" width="20" height="20" /> Cursor CLI
|
|
68
|
+
|
|
69
|
+
```bash
|
|
70
|
+
npx @vectorize-io/hindsight-coding-agents install cursor-cli
|
|
71
|
+
```
|
|
72
|
+
|
|
73
|
+
Hooks in `~/.cursor/hooks.json`, `~/.cursor/mcp.json`, and the companion skill.
|
|
74
|
+
|
|
75
|
+
#### <img src="https://hindsight.vectorize.io/img/harness/copilot-cli.svg" alt="" width="20" height="20" /> GitHub Copilot CLI
|
|
76
|
+
|
|
77
|
+
```bash
|
|
78
|
+
npx @vectorize-io/hindsight-coding-agents install copilot-cli
|
|
79
|
+
```
|
|
80
|
+
|
|
81
|
+
`~/.copilot/hooks/`, `mcp-config.json`, and the companion skill.
|
|
82
|
+
|
|
83
|
+
#### <img src="https://hindsight.vectorize.io/img/harness/grok-build.svg" alt="" width="20" height="20" /> Grok Build
|
|
84
|
+
|
|
85
|
+
```bash
|
|
86
|
+
npx @vectorize-io/hindsight-coding-agents install grok-build
|
|
87
|
+
```
|
|
88
|
+
|
|
89
|
+
Native hooks and MCP in `~/.grok/config.toml`, plus the companion skill.
|
|
90
|
+
|
|
91
|
+
#### <img src="https://hindsight.vectorize.io/img/harness/antigravity-cli.png" alt="" width="20" height="20" /> Antigravity CLI
|
|
92
|
+
|
|
93
|
+
```bash
|
|
94
|
+
npx @vectorize-io/hindsight-coding-agents install agy
|
|
95
|
+
```
|
|
96
|
+
|
|
97
|
+
Lifecycle hooks, MCP, and the `Hindsight · <bank>` status line.
|
|
98
|
+
|
|
99
|
+
#### <img src="https://hindsight.vectorize.io/img/harness/devin-cli.svg" alt="" width="20" height="20" /> Devin CLI
|
|
100
|
+
|
|
101
|
+
```bash
|
|
102
|
+
npx @vectorize-io/hindsight-coding-agents install devin-cli
|
|
103
|
+
```
|
|
104
|
+
|
|
105
|
+
Hooks in `~/.config/devin/config.json` plus MCP. Needs Node 22.5+ — see below.
|
|
106
|
+
|
|
107
|
+
#### <img src="https://hindsight.vectorize.io/img/harness/cline-cli.svg" alt="" width="20" height="20" /> Cline CLI
|
|
108
|
+
|
|
109
|
+
```bash
|
|
110
|
+
npx @vectorize-io/hindsight-coding-agents install cline-cli
|
|
111
|
+
```
|
|
112
|
+
|
|
113
|
+
A native plugin via `cline plugin install`, plus MCP and the companion skill.
|
|
114
|
+
|
|
115
|
+
Uninstall the same way: `npx @vectorize-io/hindsight-coding-agents uninstall claude-code` (or `uninstall all`).
|
|
50
116
|
|
|
51
117
|
**Devin CLI needs Node 22.5 or newer.** Its hooks pass only a session id — the conversation itself
|
|
52
118
|
lives in `~/.local/share/devin/cli/sessions.db` — so reading it depends on Node's built-in
|
|
53
119
|
`node:sqlite`. Installing `devin-cli` checks for this first and refuses (with the reason) rather
|
|
54
120
|
than wiring hooks that could never retain anything. Every other agent works on any supported Node.
|
|
55
121
|
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
new version; re-run `install` (idempotent) only when a release note says the wiring changed.
|
|
122
|
+
`install` copies what it needs into `~/.hindsight/coding-agents` and points each agent's wiring
|
|
123
|
+
there, so nothing depends on where you ran it from. **Updating** is the same command again — it
|
|
124
|
+
re-copies the runtime in place, leaving the wiring valid and every new session on the new version.
|
|
60
125
|
|
|
61
126
|
`install` merges the native wiring (hooks + MCP registration where the host wants them) into each
|
|
62
127
|
agent's own config, preserving everything already there; it is idempotent (re-run after moving the
|
|
@@ -73,7 +138,7 @@ can ask the agent itself. Manual wiring per harness, if you prefer:
|
|
|
73
138
|
```
|
|
74
139
|
|
|
75
140
|
**Claude Code** and **Codex** get their full three-hook + MCP wiring from this package's own
|
|
76
|
-
installer — `hindsight-coding-agents install claude-code` / `install codex`. This package's `bin`
|
|
141
|
+
installer — `npx @vectorize-io/hindsight-coding-agents install claude-code` / `install codex`. This package's `bin`
|
|
77
142
|
entries (`hindsight-claude-hook`, `hindsight-codex-hook`,
|
|
78
143
|
`hindsight-cursor-hook`) are the individual injection-only `UserPromptSubmit` entrypoints for a
|
|
79
144
|
minimal, hand-wired setup.
|
|
@@ -98,7 +163,7 @@ that shipped a user config — Cursor CLI, Copilot CLI, opencode and Cline have
|
|
|
98
163
|
|
|
99
164
|
```bash
|
|
100
165
|
cd /path/to/your/repo
|
|
101
|
-
hindsight-coding-agents install claude-code --import-conversations # or: install codex --import-conversations
|
|
166
|
+
npx @vectorize-io/hindsight-coding-agents install claude-code --import-conversations # or: install codex --import-conversations
|
|
102
167
|
```
|
|
103
168
|
|
|
104
169
|
This re-extracts the transcripts the agent already wrote, so it costs tokens roughly in proportion
|
|
@@ -141,9 +206,9 @@ Three modes, chosen once when you install (`install` asks on a terminal; pass `-
|
|
|
141
206
|
| `daemon` | a local `hindsight-embed` on this machine | `uv` on PATH + an LLM key for extraction |
|
|
142
207
|
|
|
143
208
|
```bash
|
|
144
|
-
hindsight-coding-agents install claude-code --server daemon
|
|
145
|
-
hindsight-coding-agents install claude-code --server self-hosted --api-url http://localhost:8888
|
|
146
|
-
hindsight-coding-agents install claude-code --server cloud --api-token <token>
|
|
209
|
+
npx @vectorize-io/hindsight-coding-agents install claude-code --server daemon
|
|
210
|
+
npx @vectorize-io/hindsight-coding-agents install claude-code --server self-hosted --api-url http://localhost:8888
|
|
211
|
+
npx @vectorize-io/hindsight-coding-agents install claude-code --server cloud --api-token <token>
|
|
147
212
|
```
|
|
148
213
|
|
|
149
214
|
Re-running `install` never re-asks: a config that already names a server is left alone.
|
package/dist/antigravity-hook.js
CHANGED
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -145,7 +145,9 @@ function readEnvConfig(env = process.env) {
|
|
|
145
145
|
function loadConfig(opts = {}) {
|
|
146
146
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
147
147
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
148
|
-
|
|
148
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
149
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
150
|
+
return resolveConfig(raw);
|
|
149
151
|
}
|
|
150
152
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
151
153
|
"bankId",
|
|
@@ -212,7 +212,9 @@ function readEnvConfig(env = process.env) {
|
|
|
212
212
|
function loadConfig(opts = {}) {
|
|
213
213
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
214
214
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
215
|
-
|
|
215
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
216
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
217
|
+
return resolveConfig(raw);
|
|
216
218
|
}
|
|
217
219
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
218
220
|
"bankId",
|
|
@@ -296,7 +298,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
296
298
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
297
299
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
298
300
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
299
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
301
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
300
302
|
var RETAIN_STRATEGIES = {
|
|
301
303
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
302
304
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
package/dist/claude-hook.js
CHANGED
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -1365,7 +1367,7 @@ async function buildSessionStartContext(args) {
|
|
|
1365
1367
|
cold = docIds !== void 0 ? docIds.size === 0 : void 0;
|
|
1366
1368
|
gitDocIds = docIds;
|
|
1367
1369
|
if (docIds !== void 0) {
|
|
1368
|
-
startSeed(cwd, { limit: cfg.seedLimit });
|
|
1370
|
+
startSeed(cwd, { limit: cfg.seedLimit, harness });
|
|
1369
1371
|
if (docIds.size === 0) {
|
|
1370
1372
|
if (cfg.codebaseSurvey !== false) {
|
|
1371
1373
|
startSurvey(cwd, {
|
package/dist/claude-stop-hook.js
CHANGED
|
@@ -212,7 +212,9 @@ function readEnvConfig(env = process.env) {
|
|
|
212
212
|
function loadConfig(opts = {}) {
|
|
213
213
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
214
214
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
215
|
-
|
|
215
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
216
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
217
|
+
return resolveConfig(raw);
|
|
216
218
|
}
|
|
217
219
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
218
220
|
"bankId",
|
|
@@ -296,7 +298,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
296
298
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
297
299
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
298
300
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
299
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
301
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
300
302
|
var RETAIN_STRATEGIES = {
|
|
301
303
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
302
304
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
package/dist/cline.js
CHANGED
|
@@ -175,7 +175,9 @@ function readEnvConfig(env = process.env) {
|
|
|
175
175
|
function loadConfig(opts = {}) {
|
|
176
176
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
177
177
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
178
|
-
|
|
178
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
179
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
180
|
+
return resolveConfig(raw);
|
|
179
181
|
}
|
|
180
182
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
181
183
|
"bankId",
|
|
@@ -278,7 +280,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
278
280
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
279
281
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
280
282
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
281
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
283
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
282
284
|
var RETAIN_STRATEGIES = {
|
|
283
285
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
284
286
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -15945,7 +15947,7 @@ async function buildSessionStartContext(args) {
|
|
|
15945
15947
|
cold = docIds !== void 0 ? docIds.size === 0 : void 0;
|
|
15946
15948
|
gitDocIds = docIds;
|
|
15947
15949
|
if (docIds !== void 0) {
|
|
15948
|
-
startSeed(cwd, { limit: cfg.seedLimit });
|
|
15950
|
+
startSeed(cwd, { limit: cfg.seedLimit, harness });
|
|
15949
15951
|
if (docIds.size === 0) {
|
|
15950
15952
|
if (cfg.codebaseSurvey !== false) {
|
|
15951
15953
|
startSurvey(cwd, {
|
package/dist/codex-hook.js
CHANGED
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -1365,7 +1367,7 @@ async function buildSessionStartContext(args) {
|
|
|
1365
1367
|
cold = docIds !== void 0 ? docIds.size === 0 : void 0;
|
|
1366
1368
|
gitDocIds = docIds;
|
|
1367
1369
|
if (docIds !== void 0) {
|
|
1368
|
-
startSeed(cwd, { limit: cfg.seedLimit });
|
|
1370
|
+
startSeed(cwd, { limit: cfg.seedLimit, harness });
|
|
1369
1371
|
if (docIds.size === 0) {
|
|
1370
1372
|
if (cfg.codebaseSurvey !== false) {
|
|
1371
1373
|
startSurvey(cwd, {
|
package/dist/codex-stop-hook.js
CHANGED
|
@@ -212,7 +212,9 @@ function readEnvConfig(env = process.env) {
|
|
|
212
212
|
function loadConfig(opts = {}) {
|
|
213
213
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
214
214
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
215
|
-
|
|
215
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
216
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
217
|
+
return resolveConfig(raw);
|
|
216
218
|
}
|
|
217
219
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
218
220
|
"bankId",
|
|
@@ -296,7 +298,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
296
298
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
297
299
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
298
300
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
299
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
301
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
300
302
|
var RETAIN_STRATEGIES = {
|
|
301
303
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
302
304
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
package/dist/copilot-hook.js
CHANGED
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -1365,7 +1367,7 @@ async function buildSessionStartContext(args) {
|
|
|
1365
1367
|
cold = docIds !== void 0 ? docIds.size === 0 : void 0;
|
|
1366
1368
|
gitDocIds = docIds;
|
|
1367
1369
|
if (docIds !== void 0) {
|
|
1368
|
-
startSeed(cwd, { limit: cfg.seedLimit });
|
|
1370
|
+
startSeed(cwd, { limit: cfg.seedLimit, harness });
|
|
1369
1371
|
if (docIds.size === 0) {
|
|
1370
1372
|
if (cfg.codebaseSurvey !== false) {
|
|
1371
1373
|
startSurvey(cwd, {
|
|
@@ -212,7 +212,9 @@ function readEnvConfig(env = process.env) {
|
|
|
212
212
|
function loadConfig(opts = {}) {
|
|
213
213
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
214
214
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
215
|
-
|
|
215
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
216
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
217
|
+
return resolveConfig(raw);
|
|
216
218
|
}
|
|
217
219
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
218
220
|
"bankId",
|
|
@@ -296,7 +298,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
296
298
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
297
299
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
298
300
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
299
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
301
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
300
302
|
var RETAIN_STRATEGIES = {
|
|
301
303
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
302
304
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
package/dist/cursor-hook.js
CHANGED
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -1365,7 +1367,7 @@ async function buildSessionStartContext(args) {
|
|
|
1365
1367
|
cold = docIds !== void 0 ? docIds.size === 0 : void 0;
|
|
1366
1368
|
gitDocIds = docIds;
|
|
1367
1369
|
if (docIds !== void 0) {
|
|
1368
|
-
startSeed(cwd, { limit: cfg.seedLimit });
|
|
1370
|
+
startSeed(cwd, { limit: cfg.seedLimit, harness });
|
|
1369
1371
|
if (docIds.size === 0) {
|
|
1370
1372
|
if (cfg.codebaseSurvey !== false) {
|
|
1371
1373
|
startSurvey(cwd, {
|
package/dist/cursor-stop-hook.js
CHANGED
|
@@ -212,7 +212,9 @@ function readEnvConfig(env = process.env) {
|
|
|
212
212
|
function loadConfig(opts = {}) {
|
|
213
213
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
214
214
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
215
|
-
|
|
215
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
216
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
217
|
+
return resolveConfig(raw);
|
|
216
218
|
}
|
|
217
219
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
218
220
|
"bankId",
|
|
@@ -296,7 +298,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
296
298
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
297
299
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
298
300
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
299
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
301
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
300
302
|
var RETAIN_STRATEGIES = {
|
|
301
303
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
302
304
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
package/dist/daemon-start.js
CHANGED
|
@@ -439,7 +439,9 @@ function readEnvConfig(env = process.env) {
|
|
|
439
439
|
function loadConfig(opts = {}) {
|
|
440
440
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
441
441
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
442
|
-
|
|
442
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
443
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
444
|
+
return resolveConfig(raw);
|
|
443
445
|
}
|
|
444
446
|
|
|
445
447
|
// src/core/daemon.ts
|
package/dist/deepen.js
CHANGED
|
@@ -287,7 +287,9 @@ function readEnvConfig(env = process.env) {
|
|
|
287
287
|
function loadConfig(opts = {}) {
|
|
288
288
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
289
289
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
290
|
-
|
|
290
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
291
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
292
|
+
return resolveConfig(raw);
|
|
291
293
|
}
|
|
292
294
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
293
295
|
"bankId",
|
|
@@ -448,7 +450,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
448
450
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
449
451
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
450
452
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
451
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
453
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
452
454
|
var RETAIN_STRATEGIES = {
|
|
453
455
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
454
456
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
package/dist/devin-hook.js
CHANGED
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -1365,7 +1367,7 @@ async function buildSessionStartContext(args) {
|
|
|
1365
1367
|
cold = docIds !== void 0 ? docIds.size === 0 : void 0;
|
|
1366
1368
|
gitDocIds = docIds;
|
|
1367
1369
|
if (docIds !== void 0) {
|
|
1368
|
-
startSeed(cwd, { limit: cfg.seedLimit });
|
|
1370
|
+
startSeed(cwd, { limit: cfg.seedLimit, harness });
|
|
1369
1371
|
if (docIds.size === 0) {
|
|
1370
1372
|
if (cfg.codebaseSurvey !== false) {
|
|
1371
1373
|
startSurvey(cwd, {
|
package/dist/devin-stop-hook.js
CHANGED
|
@@ -212,7 +212,9 @@ function readEnvConfig(env = process.env) {
|
|
|
212
212
|
function loadConfig(opts = {}) {
|
|
213
213
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
214
214
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
215
|
-
|
|
215
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
216
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
217
|
+
return resolveConfig(raw);
|
|
216
218
|
}
|
|
217
219
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
218
220
|
"bankId",
|
|
@@ -296,7 +298,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
296
298
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
297
299
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
298
300
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
299
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
301
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
300
302
|
var RETAIN_STRATEGIES = {
|
|
301
303
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
302
304
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
package/dist/grok-hook.js
CHANGED
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -238,7 +238,9 @@ function readEnvConfig(env = process.env) {
|
|
|
238
238
|
function loadConfig(opts = {}) {
|
|
239
239
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
240
240
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
241
|
-
|
|
241
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
242
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
243
|
+
return resolveConfig(raw);
|
|
242
244
|
}
|
|
243
245
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
244
246
|
"bankId",
|
|
@@ -322,7 +324,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
322
324
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
323
325
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
324
326
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
325
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
327
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
326
328
|
var RETAIN_STRATEGIES = {
|
|
327
329
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
328
330
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -1365,7 +1367,7 @@ async function buildSessionStartContext(args) {
|
|
|
1365
1367
|
cold = docIds !== void 0 ? docIds.size === 0 : void 0;
|
|
1366
1368
|
gitDocIds = docIds;
|
|
1367
1369
|
if (docIds !== void 0) {
|
|
1368
|
-
startSeed(cwd, { limit: cfg.seedLimit });
|
|
1370
|
+
startSeed(cwd, { limit: cfg.seedLimit, harness });
|
|
1369
1371
|
if (docIds.size === 0) {
|
|
1370
1372
|
if (cfg.codebaseSurvey !== false) {
|
|
1371
1373
|
startSurvey(cwd, {
|
package/dist/grok-stop-hook.js
CHANGED
|
@@ -212,7 +212,9 @@ function readEnvConfig(env = process.env) {
|
|
|
212
212
|
function loadConfig(opts = {}) {
|
|
213
213
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
214
214
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
215
|
-
|
|
215
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
216
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
217
|
+
return resolveConfig(raw);
|
|
216
218
|
}
|
|
217
219
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
218
220
|
"bankId",
|
|
@@ -296,7 +298,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
296
298
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
297
299
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
298
300
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
299
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
301
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
300
302
|
var RETAIN_STRATEGIES = {
|
|
301
303
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
302
304
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
package/dist/hindsight-seed.js
CHANGED
|
@@ -181,7 +181,9 @@ function readEnvConfig(env = process.env) {
|
|
|
181
181
|
function loadConfig(opts = {}) {
|
|
182
182
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
183
183
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
184
|
-
|
|
184
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
185
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
186
|
+
return resolveConfig(raw);
|
|
185
187
|
}
|
|
186
188
|
|
|
187
189
|
// src/core/bank.ts
|
package/dist/index.js
CHANGED
|
@@ -242,7 +242,9 @@ function readEnvConfig(env = process.env) {
|
|
|
242
242
|
function loadConfig(opts = {}) {
|
|
243
243
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
244
244
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
245
|
-
|
|
245
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
246
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
247
|
+
return resolveConfig(raw);
|
|
246
248
|
}
|
|
247
249
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
248
250
|
"bankId",
|
|
@@ -311,7 +313,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
311
313
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
312
314
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
313
315
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
314
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
316
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
315
317
|
var RETAIN_STRATEGIES = {
|
|
316
318
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
317
319
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -15943,7 +15945,7 @@ async function buildSessionStartContext(args) {
|
|
|
15943
15945
|
cold = docIds !== void 0 ? docIds.size === 0 : void 0;
|
|
15944
15946
|
gitDocIds = docIds;
|
|
15945
15947
|
if (docIds !== void 0) {
|
|
15946
|
-
startSeed(cwd, { limit: cfg.seedLimit });
|
|
15948
|
+
startSeed(cwd, { limit: cfg.seedLimit, harness });
|
|
15947
15949
|
if (docIds.size === 0) {
|
|
15948
15950
|
if (cfg.codebaseSurvey !== false) {
|
|
15949
15951
|
startSurvey(cwd, {
|
package/dist/kilo.js
CHANGED
|
@@ -242,7 +242,9 @@ function readEnvConfig(env = process.env) {
|
|
|
242
242
|
function loadConfig(opts = {}) {
|
|
243
243
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
244
244
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
245
|
-
|
|
245
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
246
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
247
|
+
return resolveConfig(raw);
|
|
246
248
|
}
|
|
247
249
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
248
250
|
"bankId",
|
|
@@ -311,7 +313,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
311
313
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
312
314
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
313
315
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
314
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
316
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
315
317
|
var RETAIN_STRATEGIES = {
|
|
316
318
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
317
319
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
|
@@ -15943,7 +15945,7 @@ async function buildSessionStartContext(args) {
|
|
|
15943
15945
|
cold = docIds !== void 0 ? docIds.size === 0 : void 0;
|
|
15944
15946
|
gitDocIds = docIds;
|
|
15945
15947
|
if (docIds !== void 0) {
|
|
15946
|
-
startSeed(cwd, { limit: cfg.seedLimit });
|
|
15948
|
+
startSeed(cwd, { limit: cfg.seedLimit, harness });
|
|
15947
15949
|
if (docIds.size === 0) {
|
|
15948
15950
|
if (cfg.codebaseSurvey !== false) {
|
|
15949
15951
|
startSurvey(cwd, {
|
package/dist/mcp-server.js
CHANGED
|
@@ -31102,7 +31102,9 @@ function readEnvConfig(env = process.env) {
|
|
|
31102
31102
|
function loadConfig(opts = {}) {
|
|
31103
31103
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
31104
31104
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
31105
|
-
|
|
31105
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
31106
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
31107
|
+
return resolveConfig(raw);
|
|
31106
31108
|
}
|
|
31107
31109
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
31108
31110
|
"bankId",
|
|
@@ -31205,7 +31207,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
31205
31207
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
31206
31208
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
31207
31209
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
31208
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
31210
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
31209
31211
|
var RETAIN_STRATEGIES = {
|
|
31210
31212
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
31211
31213
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
package/dist/status.js
CHANGED
|
@@ -212,7 +212,9 @@ function readEnvConfig(env = process.env) {
|
|
|
212
212
|
function loadConfig(opts = {}) {
|
|
213
213
|
const o = typeof opts === "string" ? { path: opts } : opts;
|
|
214
214
|
const withEnv = applyLayer({}, readEnvConfig(), o.harness);
|
|
215
|
-
|
|
215
|
+
const raw = applyLayer(withEnv, readRaw(o.path ?? CONFIG_PATH), o.harness);
|
|
216
|
+
if (!raw.harness && o.harness) raw.harness = o.harness;
|
|
217
|
+
return resolveConfig(raw);
|
|
216
218
|
}
|
|
217
219
|
var BANK_OVERRIDE_EXCLUDED = [
|
|
218
220
|
"bankId",
|
|
@@ -248,7 +250,7 @@ var GITLOG_MISSION = "You are ingesting an aggregated block of git commit MESSAG
|
|
|
248
250
|
var CONVERSATION_MISSION = `You are ingesting a developer conversation as a JSONL transcript (one {role, content} turn per line): the user's requests, the assistant's narration, and compact 'action' turns naming each tool use and its target (e.g. "Edit boltons/strutils.py") with no arguments or outputs. It may be a SHORT decision chat or a LONG working session \u2014 scale the facts to the substance, never to the message count. Extract the FEWEST facts that capture the OUTCOME: the settled DECISIONS and their exact rules/values (quote literals VERBATIM); concrete CHANGES to specific code entities; problems and how they were resolved; conventions or invariants established; at most one fact for a notable REJECTED alternative ('initially proposed X, changed to Y because Z'). A short decision chat usually yields 1-2 facts; a substantial working session several. CRITICAL: a conversation REVISES itself \u2014 record ONLY the FINAL state as what is in effect; a superseded proposal appears ONLY inside the rejected fact, NEVER as its own 'decided' fact; if the same setting changes several times keep only the LAST, and make unmistakably clear which choice WON. Do NOT emit one fact per message, per intermediate proposal, or per action turn. Keep issue/PR references (#123, GH-123) VERBATIM and emit each as an ENTITY. Preserve the 'REF-ID: <token>' marker verbatim in every fact. Do not invent; capture only what was actually settled.`;
|
|
249
251
|
var REFLECT_MISSION = "You are a debugging assistant with the project's past decisions in memory (git rationale and developer chats). Given a bug's SYMPTOM, find the past decision whose rationale explains the ROOT CAUSE \u2014 not one that merely shares vocabulary. Answer with the PRECISE fix: state the EXACT rule and the LITERAL values, identifiers, strings, numbers, or set members that were decided \u2014 quote them VERBATIM, never paraphrase, generalize, or omit them (give the actual decided value, not 'the project standard'). If memories CONFLICT on the same rule, the LATEST decision wins \u2014 prefer facts that explicitly amend or supersede an earlier one, state the superseded rule as no longer in effect, and never present it as the fix. Name the function/file to change and cite the REF-ID(s). If NOTHING in memory genuinely explains THIS symptom, say exactly that in one short sentence \u2014 a wrong-but-confident nearest match is worse than an honest miss; never stretch an unrelated decision to fit.";
|
|
250
252
|
var DOCUMENT_MISSION = "You are ingesting a standalone document (notes, docs, or structural findings). Extract the concrete facts, concepts, and structure it describes.";
|
|
251
|
-
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details.";
|
|
253
|
+
var OBSERVATIONS_MISSION = "Consolidate durable knowledge about THIS codebase \u2014 recurring patterns, conventions, module responsibilities, and how components relate \u2014 from the ingested commits and conversations. Favor stable structural understanding over one-off details. When a new fact contradicts or supersedes an existing observation, UPDATE that observation to reflect the current state rather than creating a sibling alongside it; note that the rule was revised and when, so the superseded version is visible as history rather than as a competing claim.";
|
|
252
254
|
var RETAIN_STRATEGIES = {
|
|
253
255
|
git: { retain_mission: GIT_MISSION, retain_extraction_mode: "verbose" },
|
|
254
256
|
// ONE big aggregated document (last N commit messages, no diffs) -> a larger chunk size so it stays
|
package/package.json
CHANGED
package/skill/SKILL.md
CHANGED
|
@@ -94,10 +94,10 @@ name, or by one `mapPathToBank` prefix over their parent directory.
|
|
|
94
94
|
## Install / update (for setting up another machine or harness)
|
|
95
95
|
|
|
96
96
|
```bash
|
|
97
|
-
|
|
98
|
-
hindsight-coding-agents install codex
|
|
99
|
-
hindsight-coding-agents uninstall
|
|
100
|
-
|
|
97
|
+
npx @vectorize-io/hindsight-coding-agents install all # every detected agent
|
|
98
|
+
npx @vectorize-io/hindsight-coding-agents install codex # or specific: opencode|claude-code|codex|antigravity-cli|cursor-cli
|
|
99
|
+
npx @vectorize-io/hindsight-coding-agents uninstall # removes exactly what install added
|
|
100
|
+
# updating is the same install command again — it re-copies the runtime in place
|
|
101
101
|
```
|
|
102
102
|
|
|
103
103
|
## Debugging
|