PR #7 ran under the AI-USAGE label and reported real numbers: 28 agent steps,
348s, 2,071,025 input / 17,303 output tokens, and zero cache reads or writes.
The model predicted ~$0.73 on Opus 5 for that tier. The measurement prices it
at $10.79 — the model was ~15x low.
Two wrong assumptions:
- Step count and per-step growth. `full` assumed 12 steps and 1,200 tokens per
tool result; the run did 28 steps averaging ~3,300. Cost is roughly quadratic
in steps, so this compounds. Tier defaults are re-derived from the measured
per-step growth rather than from guesses.
- Caching. The model defaulted to prompt caching on. The headroom/glm-5.2 path
reports 0 read / 0 write, so the stable prefix is paid at full input price on
every step. Budget with caching off until that column is nonzero.
Adds OBSERVED_RUNS as an append-only calibration anchor, an observed-runs
section in the report, and a regression test asserting the model stays within
2.5x of the measurement — so the next drift is caught by the suite rather than
by a surprising invoice.
Corrected blended figures at 350 PRs/month: ~$1,740 Opus 5, ~$1,755 GPT-5.6
Sol, ~$696 Sonnet 5, ~$348 Haiku 4.5, ~$70 GPT-5.6 Luna.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B11e8TZZxJyzHW7jj7KWUN
Skills — the primary now loads conditionally (each one is input tokens), per a
load table in pragent.md:
- attention-tiering: classify every PR trivial/lite/full/oversized BEFORE
reading anything, and cap file reads, linter runs and subagent fan-out per
tier. This is the cost governor; the other skills defer to its budget.
- linter-playbook: per-ecosystem detect-and-run commands scoped to changed
files, the never-install rule, and how to turn a diagnostic into a finding
instead of pasting tool output.
- security-lens: the inline security checklist for when @security isn't worth
delegating, built around a source -> sink test each finding must pass.
- malicious-change: hostile-PR detection — injection aimed at the reviewer,
install/CI-time hooks, obfuscated payloads, dependency confusion, logic
backdoors. Complements the runtime containment added in the previous commit:
that stops the agent being hijacked, this makes it report the attempt.
- comment-craft: how to write problem/fix/suggestion so a maintainer can act in
one read, and what to cut.
pilot/cost_model.py — prices a review against published Claude and OpenAI rates
(fetched 2026-08-18). Prompt sizes are measured from the factory files rather
than guessed; per-tier workloads come from the tiering budgets. The model is
explicit about the thing that actually dominates an agent loop: the whole
conversation is resent every step, so caching moves ~2.3x of the bill.
Blended over a 5/35/55/5 mix with caching on: ~$0.61/PR on Opus 5 or GPT-5.6
Sol, ~$0.24 on Sonnet 5 or Terra, ~$0.12 on Haiku 4.5, ~$0.02 on Luna. At 350
PRs/month that's ~$212 / ~$85 / ~$43 / ~$8.50.
Tests: 101 -> 122.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B11e8TZZxJyzHW7jj7KWUN
The reviewer runs an opencode agent with `bash: "*": allow` over a checkout of
the PR author's branch, and the pod holds a Gitea Write credential. Those two
facts had no wall between them.
Security
- _build_env now allow-lists the subprocess environment instead of inheriting
it, so PRAGENT_BOT_TOKEN and WEBHOOK_SECRET never reach the agent. This was
the live hole: a PR body or an AGENTS.md could ask the agent to `curl` the
token out, and it had both the value and the tool.
- sanitize_workdir deletes author-controlled agent-instruction files from the
checkout before opencode starts (AGENTS.md at any depth, CLAUDE.md,
.cursorrules, a repo opencode.json/.opencode, copilot-instructions.md).
opencode loads nested AGENTS.md as instructions, so a PR could otherwise ship
its own system prompt. They are still reviewed, as data.
- The brief fences PR title/body and diff in --- UNTRUSTED --- markers under a
trust-boundary preamble; the pragent agent, the three lens subagents and the
review-methodology skill now treat injection attempts as a critical finding
to report rather than an instruction to obey.
- .pr-review.json is read from the PR's base branch, not the head sha. Its
`instructions` field is spliced into the reviewer's prompt, so head-ref
reading let any author rewrite the reviewer's rules. Fields are length-capped.
- Untar rejects escaping symlinks, parent traversal, and writes through a
planted symlink (tar-slip).
- The image runs as uid 10001 instead of root.
Robustness
- Bounded review concurrency (PRAGENT_MAX_CONCURRENT_REVIEWS, default 2). Each
review forks an opencode process; a thread per delivery was a fork bomb on a
burst of labels or Gitea retries.
- An in-flight (repo, index, sha) claim closes the check-then-act race in the
sha-marker dedupe, where two deliveries a second apart both read "not yet
reviewed" and both posted.
- Request bodies are capped before being read into memory.
Correctness
- parse_diff_anchors counts a whitespace-stripped blank context line. Skipping
it desynced the new-line counter for the rest of the hunk and silently
misplaced every later inline comment in that file.
- post_inline_review's body-only fallback folds the anchored findings into the
body. It previously posted a summary saying "N inline comment(s) below" with
no comments and no findings — losing them all on the one path that matters.
- fetch_pr_diff's files-endpoint fallback emits real a// b/ prefixes (so
changed_files and the anchor parser work on it) and reports both HTTP statuses
in its error instead of the same one twice.
- The CI workflow template pins PRAGENT_ENGINE=ollama; review_pr defaults to
opencode, which does not exist on a Gitea Actions runner.
Tests: 68 -> 101, covering each of the above.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B11e8TZZxJyzHW7jj7KWUN
Add per-review + per-comment token accounting, surfaced only when a PR carries
the new AI-USAGE label (on top of the existing AI-REVIEW trigger).
opencode_review:
- run_opencode now uses `--format json`; parse_opencode_events reconstructs the
assistant text from `text` events and sums tokens/cost/steps from every
`step_finish` event (tolerant of noise / missing fields).
- run() measures duration_s around the opencode call and returns (text, usage).
- changed_files(diff) extracts the `+++ b/` paths; the brief now lists them
under a "Changed files" focus block so the agent grounds findings in the
diff's neighbourhood instead of unbounded whole-repo walks.
ai_review:
- format_usage_section renders a `## AI usage` block: measured totals
(in/out/reasoning/cache/cost/steps/duration), the whole-repo scope note, and
an attributed per-finding table. Per-comment counts are output tokens split by
each finding's body weight — labelled "attributed" since one model pass
produces all findings.
- inline_comment_body appends `🪙 ~N tok (X% · attributed output)` when
attribution is present.
- review_pr gains report_usage; compute_attribution stashes _tok_attrib/_tok_pct.
- format_review_body inserts the usage section between summary and findings.
webhook_server:
- Fire on every pull_request action except `closed` (denylist, was an allowlist)
— the AI-REVIEW gate + sha dedupe keep this safe.
- AI-USAGE label detection + PRAGENT_USAGE_ALWAYS env drive report_usage.
.opencode factory + review-methodology skill: new "Ground findings in context"
step — read callers/imports/sibling functions per changed file (1-3 files per
finding), no unbounded walks.
Tests: parse_opencode_events (text+usage sum, malformed tolerance, none-usage),
changed_files, compute_attribution math, inline 🪙 line, format_usage_section
totals/table/cost, format_review_body ordering. 68 passing.
Co-Authored-By: Claude <noreply@anthropic.com>
Gitea 1.26.x has no GitHub-style 'Apply suggestion' button — a ```suggestion
fence is just an unknown-language code block, so chroma does not highlight it
and there is no apply control. Switch inline_comment_body to wrap the suggested
fix in a fence tagged with the file's language (new _lang_for_path helper,
.java→java, .ts→typescript, .py→python, ...), so Gitea syntax-highlights the
code. No capability lost (there was never an apply button on this Gitea
version). Correct the docstrings/skills/README that wrongly claimed an
apply-button was rendered.
Co-Authored-By: Claude <noreply@anthropic.com>
Replace the single Python model-call reviewer with an opencode agent
factory. A primary 'pragent' agent reads a brief (title/body/diff/config/
prior reviews), inspects the checked-out repo, runs the repo's own linters
via bash, loads review-methodology + findings-schema skills, and emits a
{summary, findings} JSON with per-finding severity/path/line/problem/fix/
suggestion/reference. Dormant security/tests/perf subagent lenses fan out
only on large/risky diffs (lean by default).
pilot/opencode_review.py: fetches the repo archive at the head sha into a
temp workdir, writes .pragent/brief.md, drops the factory, runs
'opencode run --pure --agent pragent --dir <workdir>' headlessly. Isolates
HOME (shared, warmed), strips ANTHROPIC_* env (leaked host vars caused
ProviderModelNotFoundError), stdin=DEVNULL (opencode blocks on stdin),
maps the bare OLLAMA_MODEL to the provider-prefixed ref. No Gitea I/O —
ai_review.review_pr parses + anchors + posts (reuses all v2 logic/tests).
PRAGENT_ENGINE=opencode (default) selects it; =ollama keeps the legacy
direct-call path. Verified end-to-end: posts a real review with a summary
section, inline [CRITICAL]/[HIGH] comments + apply-able suggestions +
reference links, and the sha dedupe marker. 49 tests pass.
Co-Authored-By: Claude <noreply@anthropic.com>
Ships a working pragent pilot ahead of the framework build (design doc
deferred). Single stdlib-only reviewer script fetched at runtime by a per-repo
Gitea Action; reviews fire only on PRs with the AI-REVIEW label; model is
glm-5.2:cloud via the on-network headroom proxy; fail-open.
- pilot/ai_review.py: fetch PR diff, call model, post review as pragent-bot
- pilot/workflow-template.yml: per-repo Gitea Action gated on AI-REVIEW label
- pilot/README.md: onboarding (bot collaborator + secret + workflow + label)
- tests/pilot/test_ai_review.py: 15 unit tests for pure helpers (no network)
- README/design doc: note pilot is the bootstrap; framework build deferred
Co-Authored-By: Claude <noreply@anthropic.com>
Marcos