mirror of
https://github.com/pewdiepie-archdaemon/odysseus.git
synced 2026-10-07 23:42:21 +02:00
merge: reconcile Wave 1.1 with post-PR40 lab
Merge canonical lab 9557b8d5909eb4a885c3bf49e19a65dd904f8c1d exactly once. Retain invocation journal ownership and lineage, provider terminal ordering, teacher handoff, framed DONE handling, and canonical authority/Ajax routing. Combine dynamic dispatch receipts with lab policy forwarding. Adapt native shell/patch evidence, explicit TUI verifiers, and artifact recovery presentation. Refresh generated configuration source links and strengthen adapter regressions. Validation: focused 2118 passed; Wave 1.1 script 2291 passed; broad runtime 5649 passed; full pytest 11581 passed, 53 skipped, 2 xfailed, 6 subtests passed. Compileall 1689 Python files; syntax 279 JS and 82 MJS files; diff and conflict-marker checks passed.
This commit is contained in:
Alexandre Teixeira
@@ -0,0 +1,89 @@
|
||||
# Ref parity audit
|
||||
|
||||
`scripts/ref_parity_audit.py` reports which commits on one git ref left no trace
|
||||
in another, and which files exist on one and not the other. It is read-only: it
|
||||
runs `git log`, `git show`, `git diff`, `git grep`, `git ls-tree` and
|
||||
`git merge-base`, writes nothing to the repository, touches no remote, and does
|
||||
not import the application package.
|
||||
|
||||
## Why it exists
|
||||
|
||||
`lab` and the public `dev` line share only the repository's first commit as a
|
||||
merge base, so `git log lab..dev` lists thousands of commits — nearly all of
|
||||
which are in fact present on both sides, having arrived under different SHAs. A
|
||||
plain log tells you nothing about what is actually missing.
|
||||
|
||||
The question that matters before `lab` becomes a release is narrower: is there a
|
||||
fix on the public line that never reached `lab`? This script answers that by
|
||||
sampling distinctive added lines from each commit and searching the other tree
|
||||
for them.
|
||||
|
||||
## Running it
|
||||
|
||||
```bash
|
||||
git remote add public https://github.com/odysseus-dev/odysseus.git # once
|
||||
git fetch public dev --no-tags
|
||||
|
||||
scripts/ref_parity_audit.py --source public/dev --target lab --since 2026-08-10
|
||||
```
|
||||
|
||||
Roughly 30 seconds for a 100-commit window; it grows linearly, so bound a wide
|
||||
audit with `--since`. Add `--format json` for a machine-readable report and
|
||||
`--output PATH` to write it to a file.
|
||||
|
||||
| Flag | Effect |
|
||||
|---|---|
|
||||
| `--source REF` | The ref whose commits are audited. Required. |
|
||||
| `--target REF` | The ref searched for traces of them. Required. |
|
||||
| `--since` / `--until` | Bound the commit range. Both filter **committer** date, which is also the date the report prints. |
|
||||
| `--traversal linear` | Default. Individual authored commits, merges dropped. Finds a fix that arrived on a side branch. |
|
||||
| `--traversal first-parent` | One row per merge into the source branch, which reads as one row per merged pull request. |
|
||||
| `--probes N` | Probe lines sampled per commit, default 4. |
|
||||
| `--exclude GLOB` | Extra path glob whose lines are not used as probes. Repeatable. |
|
||||
| `--no-default-excludes` | Drop the built-in vendored / lockfile / binary exclusions. |
|
||||
| `--top N` | Rows shown per file list, default 50. |
|
||||
| `--repo PATH` | Repository to run in. Defaults to this checkout. |
|
||||
|
||||
## How a verdict is reached
|
||||
|
||||
For each commit in `target..source`, the script takes the patch with no context
|
||||
lines, collects the added lines, drops the ones from vendored code, committed
|
||||
build output, lockfiles and binaries, and keeps those that are at least 24
|
||||
characters long and name at least two distinct identifiers. It ranks what is
|
||||
left by how many distinct identifiers each line carries (length breaks ties),
|
||||
takes the top `--probes`, and searches the whole target tree for each one with
|
||||
`git grep --fixed-strings`.
|
||||
|
||||
Probes are stripped of leading and trailing whitespace, so a change that was
|
||||
re-indented on the target still counts as present. The whole target tree is
|
||||
searched, not the same file, because a ported fix routinely moves.
|
||||
|
||||
| Verdict | Meaning |
|
||||
|---|---|
|
||||
| **absent** | No probe found anywhere in the target. Treat as a real gap and read the diff. |
|
||||
| **partial** | Some probes found. **Inconclusive.** A line can be rewritten by a refactor on the target and still be the same change. |
|
||||
| **present** | Every probe found. The change is almost certainly there in some form. |
|
||||
| **no-probe** | Nothing to sample: a deletion-only commit, or one touching only excluded paths. No verdict. |
|
||||
|
||||
## What is exact and what is a heuristic
|
||||
|
||||
**Exact:** the two file-presence lists. They come from `git ls-tree` on both
|
||||
refs, so a file in "on the source and not the target" is definitely not there.
|
||||
|
||||
**Heuristic:** every commit verdict. It samples at most four lines out of a
|
||||
diff that may be hundreds, and a probe can be absent because the area was
|
||||
refactored rather than because the change was never made.
|
||||
|
||||
The two complement each other in a specific and useful way. A commit that reads
|
||||
**present** while one of the files it added shows up in the source-only list is
|
||||
almost always a fix whose production change was reproduced on the target without
|
||||
its test. The line sampling cannot see that; the presence diff can.
|
||||
|
||||
Read the diff before porting anything. The verdicts say where to look, not what
|
||||
to do.
|
||||
|
||||
## Tests
|
||||
|
||||
`tests/test_ref_parity_audit.py`. The end-to-end cases build a throwaway
|
||||
repository with two branches off one root, so the verdicts come from git's own
|
||||
`grep` and `diff` rather than from a fake.
|
||||
@@ -0,0 +1,153 @@
|
||||
# Wave 1.1 final post-PR40 reconciliation
|
||||
|
||||
This is the one-time local reconciliation of completed Wave 1.1 with the
|
||||
authoritative post-PR40 lab commit. It does not start another runtime wave.
|
||||
|
||||
## Verified starting state
|
||||
|
||||
- Wave branch: `feature/agent-runtime-wave-1-1`.
|
||||
- Original Wave HEAD: `63457367aeed431b2c48967988259e5861f19916`, clean.
|
||||
- Canonical branch: `lab`.
|
||||
- Canonical HEAD: `9557b8d5909eb4a885c3bf49e19a65dd904f8c1d`, clean.
|
||||
- Merge base: `f0761641a12b63e401960f596d3d1be8fc90fbea`.
|
||||
- Divergence: 10 Wave-only commits and 47 lab-only commits.
|
||||
- Changed-file overlap: `src/agent_loop.py`, `src/tool_execution.py`,
|
||||
`tests/test_tool_policy.py`, and `tests/README.md`.
|
||||
|
||||
The Wave-only commits were `d57d5c58`, `dfeab64a`, `ae2445d6`, `7d84f3fe`,
|
||||
`1470dbb2`, `32830918`, `ba29afb9`, `bdfcbc0a`, `70cbaf81`, and `63457367`.
|
||||
Their completed behavior is retained. The canonical worktree is read-only;
|
||||
the exact canonical SHA was merged once with `--no-ff --no-commit`.
|
||||
|
||||
## Semantic integration
|
||||
|
||||
The only textual conflict was in `src/tool_execution.py`, where Wave 1.1
|
||||
wrapped dynamic dispatch with `dispatched(...)` and lab added `disabled_tools`
|
||||
and `tool_policy` forwarding. The resolution retains both inside the wrapper.
|
||||
Lab's new owner-aware image-generation dispatch also receives that wrapper.
|
||||
The image regression checks that explicit denial never invokes the backend,
|
||||
actual dispatch has an execution identity, and a backend without an explicit
|
||||
exit code does not manufacture an authoritative success receipt.
|
||||
|
||||
Broad validation exposed narrow adapter incompatibilities beyond the textual
|
||||
conflict. Native host-shell JSON now uses the same decoded command classification
|
||||
as journal evidence. The exact existing TUI interpreter-selection string is
|
||||
shared with the evidence parser: a following foreground verifier keeps its
|
||||
exit status, while generic conditional discovery, help/collection modes,
|
||||
variable arguments, and status-masking tails remain insufficient test proof.
|
||||
The generated interpreter-selection command itself is unchanged.
|
||||
|
||||
The generated environment reference is refreshed with the canonical generator
|
||||
so its source-location links match the reconciled code.
|
||||
|
||||
Structured native patch arguments retain artifact targets. A pre-edit
|
||||
inspection cannot invalidate a later passing executable verifier, but still
|
||||
cannot verify the edited artifact by itself; failed post-edit inspections
|
||||
remain failures. Artifact recovery's terminal round-text revisions retract buffered rejected drafts
|
||||
before presentation; their replacement prose is still gated by journal
|
||||
evidence. Explicit final-response events retain precedence, safe reasoning
|
||||
survives, and provider-error partials and diagnostics retain their ordering.
|
||||
|
||||
Existing subprocess doubles now carry PIDs. Execution simulations use the
|
||||
existing receipt-aware test helper. Contract tests assert the additional
|
||||
completion-decision event and retain their no-inference/no-execution spies.
|
||||
TUI tests retain tool order, retry behavior, and positive explicit-verifier
|
||||
coverage while additionally rejecting completion from an opaque fallback.
|
||||
Round-control fixtures explicitly fail unconfigured direct-provider synthesis
|
||||
instead of contacting their fake endpoint, and supply the synthetic context
|
||||
window while retaining real compaction logic. Conversational round provenance is
|
||||
preserved outside artifact recovery.
|
||||
|
||||
The agent-loop changes merged automatically: lab's weather relevance and
|
||||
policy-gated browser fallback coexist with Wave's action receipts, completion
|
||||
gate, and deferred teacher handoff. The fallback dispatcher runs inside the
|
||||
current invocation's journal. No generic tool floor was restored.
|
||||
|
||||
`src/agent_runs.py`, `routes/chat_routes.py`, `static/js/chat.js`,
|
||||
`static/js/chatRenderer.js`, `src/tool_policy.py`, `src/tool_capabilities.py`,
|
||||
`src/turn_contract.py`, `src/model_profiles.py`, and
|
||||
`src/clean_agent_preview.py` retain the exact canonical lab content.
|
||||
|
||||
## Identity audit
|
||||
|
||||
These classifications describe every relevant identity use across the
|
||||
detached-run manager, chat routes/browser consumers, completion gate, journal,
|
||||
teacher handoff, and existing server-owned security provenance.
|
||||
|
||||
| Class | Uses and boundary |
|
||||
| --- | --- |
|
||||
| 1. Live/detached stream-run identity | `agent_runs._Run.run_id`, `get_run_id`, and the chat response's `X-Odysseus-Run-Id` identify the detached stream. The browser's `_streamRunIds` is populated from the response header. |
|
||||
| 2. Stop/resume/replay identity | `expected_run_id` in `stop` and `request_finish`, route request headers, `_postExactStop`, the finish-editor request, `streamRunId`, and `resumeRunId` refer to that same detached stream. `subscribe` binds the exact `_Run` object returned by start/resume. |
|
||||
| 3. Stream metrics/cost identity | `_metricsCostRecordId` uses the header-derived stream ID plus `primary`/`teacher`; `metrics._costRecordId` and the cost renderer's local `runId` refer to this accounting key. Neither uses terminal metadata's journal `run_id`. |
|
||||
| 4. Logical nested invocation identity | `ActionJournal.run_id` is generated per completion-gated invocation. `action_id` is derived from it. The completion gate's terminal metadata `run_id` identifies this logical invocation. Existing `ToolRunSecurityContext.run_id` and `origin_run_id` values identify separate server-owned invocation/skill provenance operations; they are neither stream IDs nor journal lineage. |
|
||||
| 5. ActionJournal parent/child identity | `ActionJournal.parent_run_id`, the gate's parent lookup, `_parent_run_id`, `request_teacher_takeover`'s captured parent ID, and `run_teacher_inline(parent_run_id=...)` link journal invocations. The completion metadata's `parent_run_id` preserves that lineage. |
|
||||
|
||||
No invocation ID is passed to stream stop/finish/replay APIs. No stream ID is
|
||||
inserted into journal lineage. A new detached-stream regression creates nested
|
||||
gates, rejects both journal IDs at stop/finish, accepts the stream ID for finish,
|
||||
and verifies identical replay and unchanged journal metadata.
|
||||
|
||||
## Runtime invariants and final lab behavior
|
||||
|
||||
Every gated invocation creates a distinct journal, including children using
|
||||
the same workspace. Journal and action bindings restore on normal unwind,
|
||||
exception, cancellation, and generator close. Child awaiting/exhausted/error
|
||||
state cannot rewrite the parent's completion decision or receipts.
|
||||
|
||||
The teacher adapter runs after the student gate closes. It forwards the parent
|
||||
turn contract, tool policy, disabled tools, plan, client runtime context, and
|
||||
external-untrusted-context restriction. Teacher execution receives a new
|
||||
journal whose parent is the student invocation. Inner terminal frames are
|
||||
consumed; only the outer adapter emits final termination. Exact framed
|
||||
`data: [DONE]` events are distinguished from ordinary content containing the
|
||||
literal marker.
|
||||
|
||||
Provider failures retain live events, then safe partial content when present,
|
||||
then a non-completing decision, terminal metadata, and the original error last,
|
||||
without DONE. A bare error remains a bare error. Completion gating does not
|
||||
add provider calls or turn missing evidence into extra provider rounds.
|
||||
|
||||
Lab's server-owned authority remains narrower than inventory or availability.
|
||||
Transcription, OCR, tasks, browser fallback, request-specific capability
|
||||
selection, compact contracts, and provider-compatible tool choice retain the
|
||||
canonical implementation. Model ID `Ajax` selects the Odysseus compact profile;
|
||||
its selected schema boundary survives compatible `auto` tool choice, explicit
|
||||
no-tools remains explicit, and transport remains OpenAI-compatible. No
|
||||
benchmark-runner code was independently edited or executed.
|
||||
|
||||
## Validation records
|
||||
|
||||
The current requirements were installed in an isolated environment under this
|
||||
worktree's ignored `.cache/wave1-1-reconciliation` directory. The shell's
|
||||
unrelated `python` environment was not used for the accepted validation.
|
||||
Canonical full pytest uses the repository's default data directory and allows
|
||||
dotenv loading so research-path and setup tests can exercise their own fixtures;
|
||||
the focused Wave script retains its explicit runtime isolation settings.
|
||||
Optional live Ajax tests retain their opt-in skips; no live model or benchmark
|
||||
run is part of this reconciliation.
|
||||
|
||||
- [Focused tests](validation/wave-1-1-reconciliation-focused.txt)
|
||||
- [Wave 1.1 validation script](validation/wave-1-1-reconciliation-wave-validation.txt)
|
||||
- [Broad affected runtime suite](validation/wave-1-1-reconciliation-broad.txt)
|
||||
- [Canonical full pytest](validation/wave-1-1-reconciliation-pytest.txt)
|
||||
- [Compileall, JS/MJS syntax, diff checks, and conflict-marker scan](validation/wave-1-1-reconciliation-gates.txt)
|
||||
|
||||
The focused records include the final relevant rerun after the reconciliation
|
||||
audit was written. Full pytest and canonical static gates run afterward. The
|
||||
local merge is committed only after the required checks pass. No push, PR,
|
||||
deployment, or later-wave work is authorized by this reconciliation.
|
||||
|
||||
## Maestrum limitations encountered
|
||||
|
||||
The normal read-only pre-merge comparison stalled without a completion or
|
||||
failure payload; its execution cell was terminated and the investigation was
|
||||
not retried. Exact-path inspection proceeded using `local_only` with
|
||||
`scope_mode="worktree"`.
|
||||
|
||||
The Context Firewall rejected an unbounded `git diff --cached --check` command
|
||||
and withheld raw log output after the inspection allowance was exhausted.
|
||||
Requests for ignored `.log` files were rejected with
|
||||
`scope_rejected: ignored_by_git`. Unignored `.txt` validation records were
|
||||
subsequently admitted by exact path. Canonical checks themselves run as
|
||||
validation operations and record their exit status in the admitted gate log.
|
||||
No epoch waiting or alternative worker mechanism was used.
|
||||
@@ -0,0 +1,7 @@
|
||||
Python compileall: 1689 tracked files; 0 failures
|
||||
JS syntax: 279 tracked files; 0 failures
|
||||
MJS syntax: 82 tracked files; 0 failures
|
||||
git diff --check: exit 0
|
||||
git diff --cached --check: exit 0
|
||||
git diff HEAD --check: exit 0
|
||||
Conflict-marker scan: 2377 tracked files; 0 matches
|
||||
Reference in New Issue
Block a user