Repository navigation
Releases
v1.5.1 — security + observability + ops prep for v2.0.0
Cumulative release covering everything since v1.3.0. The big themes:
Security (v1.3.1 + v1.4.0a + v1.x-scope-reduction):
- HKDF key scaffolding + 4-CVE dep floor bumps
- TrustedHost middleware + body-size cap + control-char title validation
_FIELD_VALIDATORSfor user_store withapps_script_urlvalidator (strictscript.google.comonly)- Default OAuth scopes reduced from 7 → 5 (script.* requested incrementally)
Testing (v1.4.0b + v1.4.0c):
- 8 integration tests + chaos harness (
tests/chaos/run_chaos.py) - e2e GitHub Actions workflow with 4 jobs (
.github/workflows/e2e.yml)
Observability + v2.0b prep (v1.5.0 + v1.5.1):
- Shim-hit counter + total-call denominator per purpose
- Env-var overrides for individual derived keys (
MCP_API_BEARER_KEY, etc.) - Preflight script
scripts/preflight_strict_flip.sh - RUNBOOK §3.5 (post-flip recovery) + §3.6 (preflight procedure)
Migration + LLM-UX (v2.0a + v2.2a + v2.2b + v2.3):
scripts/migrate_existing_users.py(dry-run default,--applyrequired for writes,--forceskips heartbeat)- 4 audit-produced docs: THREAT_MODEL, RUNBOOK, TOOL_CONTRACT, CONTRIBUTING
- LLM_RECOVERY artifacts: doc +
@mcp.resource("gdocs://error-recovery")+gdocs_helptool gdocs_admin_auditforensic tool (gated behindMCP_ADMIN_TOKEN)
Dependency floor bumps (deps batch via dependabot):
- fastmcp ≥3.3.1, typing-extensions ≥4.15.0, pytest ≥9.0.3, markdown-it-py ≥4.2.0, google-auth ≥2.53.0
- uv.lock regenerated
See CHANGELOG.md for full per-version detail.
Next: v2.0.0 (HKDF strict-flip) ships after this is deployed + soaked (~1h30m–2h per the in-flight TTL ceiling).
🤖 Generated with Claude Code
v1.3.0 — self-documenting MCP
The MCP now carries its own documentation. The external google-docs-fly_MCP_Reference.md becomes redundant by design.
What changed
Connect-time orientation. _SERVER_INSTRUCTIONS now includes:
- One sentence on what the server does
- 5 named workflows (goal → tool sequence):
new_doc,convert_doc_with_headings,retrofit_styled_doc,convert_sandbox_docx,cleanup - 5 operating rules (never rebuild styled .docx;
docx_pathdoesn't work from cloud chat;placeholder_behavior="rename"preserves a title page; trash tools only act on own files; first use needs interactive OAuth consent) - Pointer to
gdocs_guide/gdocs_server_info
New gdocs_guide tool. Zero-arg, always callable. Returns the same orientation as a structured payload:
{server, workflows[], operating_rules[], tool_groups{}}
Rationale: server instructions is seen only at connect time and some clients truncate or ignore it. gdocs_guide is the "start here" / --help entry point.
Workflow-aware tool descriptions. Every tool description ends with a Choreography: line (what typically comes before / after) and a NOTE: block for known failure modes — so an agent doesn't have to discover them by hitting errors:
gdocs_preview_tab_split/gdocs_tab_existing_doc/gdocs_get_signed_upload_url—docx_pathdoesn't work from cloud chatgdocs_trash_file/gdocs_untrash_file/gdocs_move_to_folder— only works on files this app createdgdocs_setup_apps_script— first call returnsneeds_authorizationwith a URL the user must open
Tests
EXPECTED_TOOLSaddsgdocs_guide(22 total);no_arg_toolsextended.- New
test_gdocs_guide_shape_includes_all_5_workflows_and_rulesasserts the 5 named workflows, 5 operating-rule topics, 7 tool_group buckets. - 240/240 unit pass; 8/8 mutations caught cleanly.
Acceptance
A fresh agent with only (a) the server instructions and (b) the tool list — no external reference file — can correctly choose and sequence tools for all 5 core workflows and avoid the known failure modes without first triggering them. gdocs_guide is the callable fallback.
v1.2.3 — hot-fix the v1.2.2 missing-fields bug
Hot-fix for v1.2.2: the live runtime's _read_mutation_check() was still returning the v1.2.1 four-field shape (stale_patches / imprecise_patches absent from gdocs_server_info), making v1.2.2's headline acceptance criterion non-verifiable from cloud chat.
Root cause
scripts/mutation_check.py's revert() used git checkout -- <file> to restore mutated sources. That works against a clean working tree but wipes uncommitted edits in any file that mutation_check also mutates. Three of the eight mutations target server.py. The v1.2.2 edit to _read_mutation_check was applied locally, then mutation_check.py ran, mutated server.py for test_trash_file_id, and reverted via git checkout — silently restoring server.py to HEAD and wiping the new return shape. The commit captured the wiped state; CI built that; live runtime served the old shape.
Fix
_read_mutation_check()now actually returnsstale_patches+imprecise_patches(v1.2.2 intent, restored).apply_mutation()returns the original file bytes on success (wasbool).revert()writes those bytes back from memory — never touches git. Uncommitted edits in mutated source files now survivemutation_check.pyruns. Workflow-safe against unclean working trees.
Tests
Two new tests:
test_revert_restores_original_bytes_not_via_gittest_revert_is_noop_when_original_is_none
19/19 unit tests pass. Local mutation_check.py: 8/8 caught cleanly.
v1.2.2 — guard against mutation-patch rot
Preventive maintenance for the mutation gate itself: detect when an injection patch has rotted instead of silently reporting "caught" for a bug that was never actually introduced.
What
scripts/mutation_check.py now classifies each mutation into one of four outcomes (was binary caught/asleep):
| Outcome | Meaning |
|---|---|
caught |
Patch applied, only the targeted test failed (clean catch) |
stale_patch |
find text gone, OR patch applied but 0 tests failed |
imprecise_patch |
Target failed AND undeclared collateral failed too |
asleep_guard |
Patch applied, but the named guard didn't notice the bug |
To distinguish, the gate now runs the full unit suite per mutation (was: targeted test only). Adds ~90s to CI's mutation job; buys collateral-damage detection.
Surfaced via gdocs_server_info
mutation_check: {
ran, caught, status, asleep_guards, # existing
stale_patches: [guard names whose patch no longer applies],
imprecise_patches: [guard names whose patch broke unrelated tests]
}
status is "passed" only when caught == ran AND all three buckets are empty. Specific subtype values (stale_patch > imprecise_patch > asleep_guard) replace the catch-all "failed".
expected_collateral
The strict "exactly one named failure" rule punished legitimate defense in depth. Mutation now accepts expected_collateral: list[str] to declare known sibling guards; failures matching those don't flag imprecise_patch. Two siblings declared on first audit (see CHANGELOG).
Tests
17 new tests in tests/unit/test_mutation_check.py for apply_mutation, _matches_nodeid (parametrized prefix match), classify_outcome (4 outcomes + expected_collateral), and aggregate (status priority).
Acceptance verified: 8/8 mutations caught cleanly, all buckets empty.
v1.2.1 — full mutation coverage
[1.2.1] — 2026-05-18
Closes the last v1.2.0 gap: mutation_check.ran goes 3 → 8.
Every named regression guard now has an automated mutation
proving it actually catches its named bug pattern.
Added
Mutations for the 5 guards that v1.2.0 deferred. Researched in
parallel by 5 named subagents (one per guard), each VERIFIED
empirically (apply patch → run pytest → expect exit ≠ 0 → revert).
All 5 came back with working diffs; 8/8 caught on local run.
-
test_owned_by_app_agrees_with_trash_outcome— flip the
403-probe branch fromwrite_results[fid] = FalsetoTrue.
Probe lies about writability → find claimsowned_by_app=True
but trash still 403s → cross-tool inconsistency assertion fires.
Reintroduces the v0.19.0 bug pattern. -
test_inject_matches_fragmented_runs— addbreakafter
the first<w:t>text append in_extract_visible_text.
Extraction stops at the first run; fragmented paragraph
["Sec", "tion", " ", "Banner"]becomes just"Sec". Marker
"Section Banner" can't match. Reintroduces the pre-v0.15.1
text-extraction bug. -
test_preview_flags_what_convert_truncates— drift
TITLE_MAX_CHARSfrom 50 to 60 inpreview.py. The test's
fixture heading is exactly 60 chars, so60 > 60becomes False
and no warning fires; even if one fired, the message would
interpolate "60" not "50" (test asserts "50" in msg). Double
failure mechanism — robust catch. -
test_auth_pkce_consistency_every_url— the hard one from
v1.2.0. Original attempt commented outflow.code_verifier = ...
butFlow.authorization_url()auto-generates a 128-char verifier
whencode_verifier=None AND autogenerate_code_verifier=True
(lib default). PKCE survived via the fallback. New mutation
overrides BOTH paths AFTER the original assignment:
flow.code_verifier = None; flow.autogenerate_code_verifier = False.
URL emits nocode_challenge. Test's first assertion fires
immediately. -
test_tool_discoverability_via_server_info— slice[1:]
on the sorted tool list returned bygdocs_server_info. Drops
the alphabetically-first tool (gdocs_add_tabs) while leaving
mcp.list_tools()intact → set-equality fails ANDtool_count
diverges (20 vs 21).
Process note
The 5 mutations were dispatched as parallel subagents. Each had
the same brief: read the test, read the code, design a minimal
find/replace, VERIFY by running pytest on the mutated state,
revert via git checkout --. All 5 returned in ~3 minutes
elapsed (vs ~15 min serial). Each agent's verified_exit_code
field made the integration step purely mechanical — no guessing
whether a mutation would actually fire.
Tests
scripts/mutation_check.py now contains 8 mutations. CI's
mutation job runs them all in ~30s. Local run confirms
8/8 mutations caught. After this commit deploys via CI,
gdocs_server_info.test_suite.mutation_check.caught/ran reports
8/8 with asleep_guards: [].
v1.1.1 — hot-fix cycle
[1.1.1] — 2026-05-18
Post-1.1.0 hot-fixes from the first real cloud-chat user testing.
Each was caught in production by the user noticing mid-use; v1.1.1
adds named unit-test regression guards for every one so the next
cycle catches them in CI.
Fixed
-
Apps Script
deployments.createrejected theentryPointsfield.
Google's API returnsInvalid JSON payload received. Unknown name "entryPoints": Cannot find field.Web-app entry-point configuration
belongs in theappsscript.jsonmanifest, NOT the deployment body.
RemovedentryPointsfromdeploy_webapp's request body; removed
the now-unusedexecute_as/accessparameters from its signature.
Guard:test_gas_deploy.py::test_deploy_webapp_body_does_not_include_entryPoints. -
Apps Script Web App manifest changed
access: MYSELF→
ANYONE_ANONYMOUS. In single-tenant v1.0 the operator was both
deployer and runtime caller, soMYSELFworked via session magic.
In v1.1 multi-tenant cloud, the USER deploys the Web App but the
SERVER calls it — unauthenticated.MYSELFwould 401 every call.
ANYONE_ANONYMOUSis the right setting for the v1.1 architecture.
Surface is bounded by the script's logic (only acts on doc IDs in
the request, only on docs the deployer owns); v1.2 will add HMAC
request validation for defense in depth. -
OAuth callback failed on Fly with
OAuth 2 MUST utilize https.
Fly terminates TLS at the edge; inside the containerrequest.url
has schemehttp://even though the public URL is HTTPS.
oauthlib.Flow.fetch_tokenvalidates the URL and rejected any
http://. Fixed by rewriting the scheme on the URL we hand to oauthlib
whenbase_urlbegins withhttps://. Did NOT set
OAUTHLIB_INSECURE_TRANSPORT=1(that disables transport security
globally). -
PKCE handling was non-deterministic; callback failed with
Missing code verifier. Auth URLs sometimes included
code_challengeand sometimes didn't, depending on which Flow code
path generated them. v1.1.1 makes PKCE always-on: every
build_authorization_urlcall generates acode_verifiervia
secrets.token_urlsafe(48), persists it server-side keyed by the
state token's nonce (seeoauth_state._pending_verifiers), and
retrieves it on callback soFlow.fetch_tokencan complete the
exchange. Guard:
test_oauth_google.py::test_auth_pkce_consistency_every_url. -
Doc-string overpromise: tools claimed to work "without setup".
gdocs_setup_apps_script's description conflated two prerequisites:
(a) the Apps Script Web App setup itself, (b) the base Google OAuth
grant. Other tools don't need (a) but ALL tools need (b). Saying
"works without setup" unqualified misled the model into trying calls
that returnedneeds_authorization. Rewrote to distinguish the two
grant types explicitly. Guard:
test_tool_schemas.py::test_tool_descriptions_truthful. -
gdocs_reset_authorizationwas registered but undiscoverable via
tool_search. Tool was visible ingdocs_server_info.tools(count
20) but search ranker couldn't surface its schema for keywords like
"reset authorization" / "revoke grant" / "sign out". Root cause:
bland leading description sentence. Rewrote to embed the synonym
set ("reset / revoke / clear stored Google OAuth credentials. Force
re-consent.") in the first 200 chars where the ranker weighs most.
Guard:test_tool_schemas.py::test_tool_discoverability_via_server_info.
Added
-
gdocs_reset_authorizationMCP tool. Clears the user's stored
Google OAuth credentials and (optionally withfull=True) Apps
Script setup state. Forces the next tool call back into the
needs_authorizationflow. Required as a recovery path AND as the
only way to re-trigger consent for testing PKCE / scope changes /
account switches. Per-user in cloud mode (via user_store); per-
machine in stdio mode (deletes~/.google-docs-mcp/token.json). -
Version string now embeds the git commit SHA as semver build
metadata.gdocs_server_inforeportsversionas
f"{__version__}+{GIT_COMMIT}"(e.g.1.1.1+abc1234) when
GIT_COMMITenv var is set. Every deploy from a distinct commit
reports a unique version string without requiring a manual
pyproject.tomlbump on every hot-fix. Per semver §10 the build-
metadata segment is informational only and doesn't affect sort.
Tests
- ~40 new test cases (parametrized over 20 tools):
test_tool_discoverability_via_server_info— server_info.tools
matches mcp.list_tools() exactly.test_tool_descriptions_truthful(parametrized over 19 OAuth-needing
tools) — no description contains "without setup" / "without
authorization" unqualified.test_tool_input_schema_non_empty(parametrized over all 20 tools)
— every tool's schema has properties or is on the no-args allowlist.test_tab_nesting_depth_cap_enforced— 4-level nesting raises
ValueError before any Google API call.test_auth_pkce_consistency_every_url— 5 sequential calls all
return URLs with code_challenge + code_challenge_method=S256, all
with unique challenges (verifier regenerated per call).test_pkce_verifier_roundtrip(+ 2 related) — sign_state with
code_verifier → verify_state returns it on consume; single-use;
no-PKCE returns None for backward compat.
Total: ~200 unit + 4 live tests. CI gates deploys on unit pass via
deploy.sh.
Internal
- Auto version-bump-on-deploy wired via the GIT_COMMIT build arg in
deploy.sh. Every push to Fly carries a unique build identifier. - GitHub Actions runs the full unit suite across Python 3.10–3.13 on
every push/PR. deploy.shrunspytest tests/unit -qbeforeflyctl deploy;
refuses to deploy on test failure (bypassable withSKIP_TESTS=1
for emergency hot-fixes).
v1.1.0 — multi-tenant cloud auth
[1.1.0] — 2026-05-18
Multi-tenant cloud auth. The remote HTTP MCP (claude.ai connector
via Fly.io) is now genuinely multi-tenant — each cloud-chat user
operates on their own Drive with their own Google identity. Before
1.1, the entire Fly deployment was single-tenant: every cloud-chat
user's tool calls implicitly used the operator's cached OAuth token,
so docs were created in the operator's Drive and the retrofit Apps
Script Web App (deployed access: MYSELF) was unusable for anyone
but the operator.
New: gdocs_setup_apps_script MCP tool
One-shot setup for the per-user Apps Script Web App needed by
gdocs_tab_existing_doc (the lossless content-move path). Run once
per user; idempotent on retry (resumes from the last successful
step). Stdio mode keeps calling the v1.0 local CLI; HTTP mode
deploys a Web App into the calling user's Drive.
Cloud architecture: "Shape C"
- FastMCP's
GoogleProviderhandles claude.ai connector auth
(identity-onlyopenid emailscopes, butvalid_scopesadvertises
the full Workspace union so consent grants Docs/Drive/Apps Script
at the same time) - Separate auth-code flow we own (
/oauth/google/api/callback)
obtains tokens we can actually use for Google API calls - Per-user state in SQLite on the Fly volume, keyed by Google
sub - Per-user lock on refresh so two concurrent tool calls don't
rotate each other's refresh_token
Why not just rely on GoogleProvider for the API tokens too: the
upstream Google tokens it holds live in private attributes
(_upstream_token_store, no public getter, no API contract). The
MCP spec blesses two-flow designs in the "URL Mode Elicitation for
OAuth Flows" pattern; production reference is
taylorwilsdon/google_workspace_mcp.
New env vars (HTTP deploy)
Required when running the Fly server with the new auth wiring:
GOOGLE_OAUTH_BASE_URL— public HTTPS hostname of the deployment
(e.g.https://my-app.fly.dev). Must exactly match the actual URL
or claude.ai's connector OAuth discovery silently fails.GOOGLE_OAUTH_CLIENT_SECRETS_JSON— full OAuth client JSON inline,
orGOOGLE_OAUTH_CLIENT_SECRETS_PATHpointing at a file.MCP_BEARER_TOKEN(already required) — reused as the HMAC signing
key for the OAuth state parameter.
Auto-set by configure_auth_for_http:
OAUTHLIB_RELAX_TOKEN_SCOPE=1 (so partial-grant consents don't
crash the OAuth callback).
Breaking: existing claude.ai connector users must reconnect
The OAuth scope set is changing. Existing connector connections need
to be disconnected + reconnected in claude.ai's connector settings to
pick up the new scopes.
Internal modules added
user_store.py— SQLite per-user state (WAL mode, per-path init
guard, merge-semantics on save, typo-rejection)oauth_state.py— HMAC-signed, single-use state-param for OAuth
callback (CSRF + replay protection)oauth_google.py—google_auth_oauthlib.Flowsetup, callback
code-exchange, GoogleProvider activationcredentials.py—get_credentials_for_user(sub)with per-user
refresh lock,invalid_grant→NeedsReauthErrormappingsetup_apps_script_for_user— cloud-side variant of
setup_apps_script_autousinguser_storeas the per-user ledger
Production bug fix in user_store
The concurrent-writes test surfaced a real PRAGMA journal_mode=WAL
race that busy_timeout doesn't mitigate. Fixed via per-path init
guard under threading.Lock — would have fired on Fly the moment
two cloud users finished OAuth simultaneously.
Tests
+49 unit tests across the v1.1 modules:
test_user_store.py (13), test_oauth_state.py (12),
test_oauth_google.py (13), test_credentials.py (11),
test_setup_apps_script_for_user.py (9),
test_phase6_consumer_branching.py (7),
test_configure_auth_for_http.py (9). Total: 158 unit + 4 live.
Dependencies
fastmcp>=2.13(was>=2.0) —GoogleProvider+valid_scopes+
_default_scope_strpost-init patch all require 2.13+.
v1.2.0 — CI-gated deploys + mutation testing
[1.2.0] — 2026-05-18
Closes the "is the gate real?" gap. Two big additions: CI-driven
deploys (so a broken commit literally cannot reach production) and
automated mutation testing (so the suite proves it catches bugs
on every build, not just once when written).
Added
-
CI-gated deploys via GitHub Actions. New
.github/workflows/deploy.yml:- Triggers on push to
main(andworkflow_dispatch). - Jobs:
unit→mutation→deploy. Each depends on the
previous; deploy runs ONLY if both test jobs are green. - The
unitjob runspytest+ injects provenance into
test-results.json(git_commit, ci_run_url from
$GITHUB_SERVER_URL/$GITHUB_REPOSITORY/actions/runs/$GITHUB_RUN_ID,
sha256 digest of the canonicalized payload). - The
mutationjob runsscripts/mutation_check.py— see below. - The
deployjob downloads both artifacts, runsflyctl deploy
with provenance build-args, runs a /health smoke check. - Concurrency: cancels in-flight deploys when a newer commit
supersedes.
One-time setup: the
FLY_API_TOKENrepo secret must be set:flyctl tokens create deploy -a sundeepg98-docs-mcp -j | jq -r .token | gh secret set FLY_API_TOKENPipes the token directly from flyctl to gh — it never appears in
shell history or chat.After this is set, every push to main runs CI; if any test fails,
the deploy doesn't happen.ci_run_urlin the test_suite block
becomes a real GitHub Actions URL (replacing"local"for
CI-built deploys). - Triggers on push to
-
scripts/mutation_check.py— automated mutation testing.
For each named regression guard, applies a known bug-injecting
patch (e.g. revertfile_id: str | list[str]→file_id: str),
runs pytest filtered to that guard, asserts the guard goes red.
If any guard does NOT catch its mutation, the build fails — that
guard is "asleep" and can't be trusted.v1.2.0 ships mutations for 3 of the 8 named guards (the cleanest
string-replace patches):test_trash_file_id_accepts_str_or_list(file_id schema)test_deploy_webapp_body_does_not_include_entryPoints(the
v1.1.1 Apps Script bug)test_tool_descriptions_truthful(the v1.1.1 docstring lie)
Other 5 guards (
test_owned_by_app_*,test_inject_matches_*,
test_preview_*,test_auth_pkce_*,test_tool_discoverability_*)
need multi-line diffs or library-quirk workarounds and are
documented as TODOs in the script. Each one added = the gate
gets one notch sharper. Iterative ship. -
test_suite.mutation_checkblock ingdocs_server_info:mutation_check: { ran: int, caught: int, status: "passed" | "failed" | "unknown", asleep_guards: [list of guards that failed to catch their bug], }Populated from
mutation-check.jsonbaked into the image by CI.
status: "passed"requiresran > 0 AND caught == ran. Empty
artifact or local-only deploy →"unknown".
Changed
- Local
deploy.shis now the emergency fallback. Primary
deploy path is push-to-main → CI. Local./deploy.shstill
works for hot-fixes that need to bypass CI (e.g. CI itself is
broken); in that caseci_run_urlreports"local"and
mutation_check.statusreports"unknown"— those values are
the signals "this build didn't go through CI."
Tests
- extended
test_server_info_includes_test_suite_blockto assert
the newmutation_checksub-block is always present with a valid
status. Total: 212 unit + 5 live, all green.
v1.1.4 — complete named-guard coverage
[1.1.4] — 2026-05-18
Closes the two gaps surfaced by the gdocs_test_manifest audit on
v1.1.3.
Fixed
-
named_regression_guards.missingwas non-empty because two
named guards lived only intests/integration/(gated behind
--live) and so didn't appear in the deploy artifact's
test-results.json (which comes frompytest tests/unit -q).-
test_owned_by_app_agrees_with_trash_outcomeadded as a
new unit test intest_soft_failure_contracts.py. Mocks both
the write-probe (used byfind_doc_by_title) and the trash
update (used bytrash_drive_file) to share a single backing
behavior, then asserts they agree across both the app-owned and
external-file scenarios. Complements the existing live
integration test (which still runs the full real-Drive E2E
when invoked with--live). -
test_preview_flags_what_convert_truncatesmoved from
tests/integration/test_title_threshold.pyto
tests/unit/test_preview_threshold_consistency.py. The original
was mislabeled as live (it tooklive_credsas a fixture but
never used it —preview_tab_splitruns locally for the
docx_path=input mode). No live coverage lost; the test
asserts the same contract, now in CI.
After this,
gdocs_test_manifest.named_regression_guards.missing
is empty — all 8 named guards present in unit suite. -
-
ci_run_urldefaulted to""which conflated "no CI run
exists yet" with "should have been set but wasn't." Per the
v1.1.3 spec, empty must now be reserved for "broken pipeline."
Deploys not from CI now reportci_run_url: "local"explicitly.
Tests
- New unit test (5 mocks-with-batch-callback plumbing): proves the
find-probe and trash-update return-value relationship is
consistent. The mock setup is more involved than typical unit
tests because the production code uses Drive's batched-HTTP
pattern with per-request callbacks. - Moved test stays green in its new home.
- Total: 212 unit + 5 live (was 6 — test_title_threshold.py removed).
v1.1.3 — verifiable test_suite
[1.1.3] — 2026-05-18
Closes "verify the test_suite block isn't just a number to trust"
gap. Three additions that make the suite independently verifiable.
Added
-
test_suite.ci_run_url— link to the GitHub Actions run that
produced this artifact. Populated bydeploy.shvia best-effort
gh run list --commit=<sha>; empty string if no run found
(deploy ran before CI completed, gh not installed, etc.). -
test_suite.report_digest— sha256 of the canonicalized
test-results.jsonpayload (excluding the_metablock itself,
chicken-and-egg). Stored in_meta.digestin the JSON file at
deploy time; recomputed by the server at read time and compared. -
test_suite.status: "tampered"— new status value emitted
when the recomputed digest doesn't match the stored one. Catches
post-build edits to the artifact'ssummary(e.g. someone hand-
editing the passed count). The status hierarchy is now:unknown: artifact missing or summary empty (SKIP_TESTS path)tampered: stored digest doesn't match recomputed digestfailed: any test failedpassed: all green AND digest verifies
-
gdocs_test_manifest()MCP tool — surfaces the test inventory- per-test outcomes from the CI artifact. Returns:
{ status: "ok" | "unknown" | "tampered", total: int, tests: [{nodeid, outcome}, ...], named_regression_guards: {present: [...], missing: [...]}, }Lets any caller confirm specific named guards (e.g.
test_owned_by_app_agrees_with_trash_outcome) actually exist and
passed — instead of trusting an opaque "203". Tool count: 20 → 21.
Fixed
- Lazy cwd evaluation in
_find_test_results_path— was
computing candidates at module-load time, freezing the working
directory. Caught bytest_test_suite_status_tampered_when_digest_ mismatcheswhich monkeypatcheschdir. Now evaluated at each
call.
Tests
test_canonical_digest_excludes_meta_block_and_is_stable— same
payload in different dict-iteration orders → identical digest;
tampering changes the digest.test_test_suite_status_tampered_when_digest_mismatches— the
killer guard: edit summary.passed without re-signing → server
reports status="tampered".test_gdocs_test_manifest_exists_and_returns_required_shape—
manifest tool returns the documented shape regardless of artifact
presence.- All 21 tool's
test_tool_descriptions_truthfuland
test_tool_input_schema_non_emptyextended to the new tool
(gdocs_test_manifest joins no_args allowlist).
Total: 210 unit + 6 live tests, all green.
Deferred to v1.2.0
- CI mutation testing stage — automated proof that injected
regressions turn their named test red. Substantial CI workflow
changes; separate atomic commit. The manual adversarial test
(branch + PR #8) already proved the loop works on file_id; the
v1.2 work is automating that across all 8 named guards on every
build.