fix(redact): close emission gaps - env suffix keys, control-char splits, process(list) (#77484)

[thatssoheil]

Fixes #77484

Closes the three-part redaction emission class verified live on main:

1. _ENV_ASSIGN_RE misses *_KEY/*_PASS/*_PW (+ lowercase env names)

• Added bare KEY/PASS/PW suffixes to _SECRET_ENV_NAMES (FAL_KEY=…, MYSQL_PASS=…, DB_PW=… now mask)
• New _ENV_ASSIGN_LOWER_RE catches lowercase env names (openai_key=…) — underscore-boundary only, so password=/token= in prose/URLs/form bodies are untouched (URL guard: skipped when :// present)
• _key_has_secret_keyword all-caps branch now uses word-boundary matching: KEYBOARD=/PASSAGE= (prose) stay, API_KEY/FAL_KEY/DB_PW mask

2. Control-char / zero-width token splits escape every regex

sk-abc\x1bdef…, ghp_abc\n123…, zero-width splits — the contiguous _PREFIX_RE couldn't match across the split, leaking the secret verbatim. New _mask_control_split_tokens strips control chars for matching (token becomes contiguous, matching even when each fragment alone is too short), then masks the original span — with guard rails so unrelated env lines (SHELL=/bin/bash\nEXA_API_KEY=…) never get joined into a false match.

3. process(action=list) returns unredacted output

tools/process_registry.py:2495 returned raw command[:200] + output_preview[-200:] with no redaction wrap, unlike poll/log/wait/kill. The list branch now runs each entry through _redact_process_result.

Tests

• 8 new redact regression tests (TestBareSecretEnvSuffixes, TestControlCharSplitTokens) — RED on old code
• 1 new process_registry test (test_list_redacts_command_and_output) — RED on old code
• Full tests/agent/test_redact.py + tests/tools/test_process_registry.py + browser/kanban redaction suites: 144/144 green

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[alt-glitch]

This was generated by AI during triage.

Related to #77768, which already addresses the #77484 redaction and process-list gaps and also covers checkpoint persistence and masked-display control characters. The control-character handling differs, so these are competing implementations rather than duplicates; consider consolidating.

[thatssoheil]

Review resolved (dbbf447): the reviewer found 4 of 9 new tests false-passed on old code — test_lowercase_env_name_masks used a sk- value (old _PREFIX_RE masked it anyway), and the 3 control-split tests asserted the whole token absent (trivially true since the splitter fragments it in ANY output, even a full leak).

Fixed:

• lowercase-env test now uses an opaque value (openai_key=xyzzyplugh1234567890abcd) only the env-name regex can catch
• control-split tests now assert the longest fragment is absent, and use bare tokens (no KEY= context)

Re-verified RED-GREEN: on old code 5/8 new tests fail (suffix keys, lowercase env, all 3 control-splits); on new code 8/8 pass. Full 4-suite run: 144/144 green.

[kshitijk4poor]

Merged via #80965 — all four of your commits were cherry-picked with your authorship preserved (you show as author on main for each).

The salvage combined the best halves of your two PRs: this PR's body (span-local control-split masking, word-bounded all-caps matching, the hardened #77484 tests, process(list) via inline map) plus the two pieces #77768 had that this one lacked — the mask_secret control stripping (#55319/#55321) and the processes.json checkpoint redaction.

One follow-up landed on top: CI caught that _mask_control_split_tokens joined a complete token at end-of-line with the next text line (browser accessibility annotations lost "button [ref=e3]" into the mask). Fixed by skipping the join when a fragment already matches _PREFIX_RE on its own — all the smuggling shapes still mask.

Thanks for the thorough work across both iterations, including hardening your own tests after the first review round.