You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
decision_source no longer reads "default" on TLS handshake
failure, malformed-tunnel HTTP, or body-read-failure paths (#139).
Those three paths now carry distinct values: tls_handshake_fail, malformed_tunnel, and body_read_fail. default retains its
narrow meaning of "no rule matched, default mode applied."
Operator dashboards filtering for failure paths by decision_source=default must update to the new values.
Audit entries from the proxy→origin TLS path now carry tls_error_category (#138) — previously only client→proxy TLS
failures populated this field. The category distinguishes upstream_cert_invalid / upstream_connect / unknown for
operators triaging origin handshake failures.
Policy
prior_decision rule clause no longer matches prior LLM advisor
verdicts (#141). The match surface is scoped to human +
static-policy decisions; an LLM verdict that resolved a previous
request cannot be silently re-applied by a deterministic rule
without re-consulting the advisor.
Observability
audit.Logger.LevelFiltered() returns a per-process counter of
entries the audit-level filter dropped (#143 part a). Distinct from Dropped() (OverflowDrop budget-exceeded drops). Lets an operator
confirm filtering is engaged vs. silently losing entries.
On startup, when audit_level != all, the operational log emits event=audit_level_filter_active once (#143 part b) so an operator
seeing fewer audit entries than expected reads the cause inline.
CI / tooling
New CI lanes (Linux): scripts/check-model-strings.sh catches
hardcoded model identifiers outside the wizard / translator
allowlist (#155); scripts/check-doc-links.sh validates every
relative markdown link across the repo's *.md files (#151).
Both are exposed as make targets.
Optional pre-commit hook (#154): scripts/precommit-check.sh
refuses to add staged files over 5 MiB. Override per-commit with TROLLBRIDGE_LARGE_FILE_OK=1. Install via ln -s ../../scripts/precommit-check.sh .git/hooks/pre-commit.
New gated test lane: go test -tags=live_az ./cmd/trollbridge/
exercises the real az CLI's JSON shape so the wizard's
Cognitive-Services parsers stay locked to Azure's actual
responses (#148).
Telemetry
advisor_consulted and advisor_classified log lines carry a
new model attribute (#157) — the AOAI deployment name (parsed
from the endpoint URL) or the configured llm.Model for other
providers. Multi-deployment AOAI operators can now attribute
log entries by deployment. Attribute is omitted when the
advisor service has no configured ModelIdentifier (back-compat).
TUI
The approvals-pane reload-failed badge now names the failing
source (#145): ␇ config reload failed / ␇ rules reload failed
/ ␇ lists reload failed instead of the bare ␇ reload failed.
Unknown / legacy-empty source falls back to the bare badge.
Internal refactor: shared formatOpRow helper consolidates the
approvals-pane row formatting that bordered and no-border
renderers previously duplicated (#142). No visual change.
Test coverage
internal/control/control_test.go gains HTTP integration tests
for /v1/rules across the three reload-status states (#144).
Subprocess pty test for Ctrl-L hard-clear sequence (#159).
E2E test exercising trollbridge verify --json against a running
daemon (#149).
E2E test exercising the audit-init startup_failure branch (#146
partial). Remaining branches tracked in #166.
Init
init AOAI provider flow now surfaces a one-line note before
the model prompt (#158): the deployment name in the endpoint URL
drives routing; the model: field is informational for AOAI.
Release tooling
scripts/release.sh --build-only (#153). Skips clean-tree
preflight and the bump/tag/push/publish flow; runs build_matrix
against the current version so implementers can verify
matrix-output without dirtying release semantics. Mutually
exclusive with the other release-flow flags.
Docs
README opening rewritten to unify voice with trollbridge.dev: the
tagline ("Let your agents run amok — but only where you say") and
the three pillars (read/write/run · outbound only where you say ·
hold the rest for one keystroke) now mirror the site's three
feature panels. The technical "policy-governed conditions"
framing moves into the body. README also gains a badges row
(release / ci / license / go version) and a one-line summary
over the audience-routing block. No CLI behavior changed.
Release process
GitHub release bodies are now sourced from CHANGELOG.md ## Unreleased
via gh release create --notes-file rather than the prior --generate-notes which had degraded to bodies reading only **Full Changelog**: <compare-link> for v0.7.10..v0.7.14. scripts/release.sh now preflights that ## Unreleased is non-empty
and fails closed otherwise, promotes the section to a versioned
heading (## vX.Y.Z — YYYY-MM-DD), and includes CHANGELOG.md in the
release bump commit. Producer-side contract documented in AGENTS.md: operator-visible changes add their CHANGELOG entry in
the same commit, not deferred to release time.
Wire / TUI
/v1/lists and /v1/llm-digests control-plane endpoints (#99). The
attach-mode TUI's URLs and LLM panels now render against a remote
daemon. Existing consumers that did not exercise either endpoint
are unaffected.
Trollbridge-Discovery header on every 470 / 471 response (#95).
Points at http://config.trollbridge.dev/discovery, a new JSON
document describing the wire protocol (status codes, headers,
body shapes, audit-log correlation). Existing agents that ignore
unknown headers continue to work unchanged.
Operator
trollbridge update --prefix <dir> (#108 part 1). Forwards the
given directory as TROLLBRIDGE_INSTALL_DIR to install.sh so the
binary lands where the operator chooses rather than install.sh's
default. The cross-repo install.sh PATH detection is tracked
separately.
trollbridge update --check (#102 part 2). Prints the latest
released version without invoking the installer.
trollbridge update now classifies failures and prints a one-line
hint above the wrapped error (#102 part 1): network /
bash_missing / permission_denied / signature_mismatch / unknown.
init daemon-mode on Windows now emits a clear refusal instead of
POSIX commands the operator cannot execute (#101 part 2). Pick
user-mode at the install-mode prompt; daemon-mode-Windows is
tracked separately.
Hot-reload now covers mode and the approvals knobs
(timeout_seconds, signal_after_seconds, on_timeout, max_pending) in addition to lists + rules (#111). Restart-only
sections are documented in internal/server/server.go HotReloadableSections.
External-edit detection migrated from mtime polling to fsnotify
(#110). Detection latency drops from the previous 2s floor to
~50ms (the debounce window).
Config and rule files are now decoded strictly (#123). A YAML key
with no matching field — a typo (mod: for mode:), or an
unsupported block — fails the load loudly instead of being silently
discarded; trollbridge validate now reports the offending key and
line instead of OK. Operator-visible breaking change: a config
or included rule file carrying an unknown or forward-compat key that
previously loaded will now error on run, validate, and
hot-reload — run trollbridge validate to find the offending key,
then remove or correct it. On the hot-reload path the reload fails
and the prior config/rule set is kept (the daemon does not crash).
Config and rule files with more than one YAML document now fail the
load (#126). ----separated documents after the first were silently
ignored — the same silent-drop class as #123 at document
granularity. A bare trailing --- separator with no content is
still accepted. Operator-visible breaking change for anyone who
split a config or rule file into multiple documents: keep one
document per file.
trollbridge run now emits a structured config_load_failure
operational-log event when it cannot load its config or open its
operational log at startup (#128). Previously these pre-startup
failures reached stderr only, with no structured event — a daemon
that failed to start left no operational-log record of why.
Rule files containing match.tool now fail to load (#125). The
field was parsed but never evaluated — a tool: clause was a
silent no-op. With strict YAML decoding (#123) the same clause now
surfaces as a parse error naming tool, so the operator can edit
the line out instead of believing a non-functional rule was active.
trollbridge init interactive wizard now defaults the model:
prompt by provider (#131). Azure OpenAI does not serve Claude
models, but the wizard had been suggesting claude-opus-4-7 after
the operator picked aoai. New defaults: claude-opus-4-7 for
anthropic (unchanged), gpt-4o-mini for aoai, empty for other.
Operators who type the model name they actually want are
unaffected.
Windows release artifacts now ship as a bare .exe instead of a .tar.gz containing trollbridge.exe (#130). The Linux/macOS
release shape is unchanged — those still get trollbridge_v<X.Y.Z>_<os>_<arch>.tar.gz. Windows operators with
bookmarked tarball URLs from prior releases will need to point at
the new trollbridge_v<X.Y.Z>_windows_<arch>.exe asset.
trollbridge validate --json emits a single JSON object on stdout
with the same shape as the existing human summary (#127). The
exit-code contract is now documented in the command help and in
the README: 0 = valid, 1 = invalid (any reason). Operators
binding config-lint from their own CI now have a stable surface.
trollbridge run now emits a structured startup_failure
operational-log event when it fails to construct after the
operational log is open (#134). Sibling to #128, which covers the pre-opLog slice (config_load_failure). The new event covers policy.NewEngine, audit.New, audit.ParseLevel, server.NewWithLoggers, and inline-list parse failures; each
carries a stage attribute (policy / audit / audit_level / server / lists). Operators alerting on "daemon failed to
start" should extend their config_load_failure query to
include startup_failure.
trollbridge init interactive wizard now hints when az is
installed but the operator is not authenticated (#136). On the aoai provider branch, the wizard distinguishes "no az in
PATH" (still a silent skip, unchanged) from "az present but az account show fails" — the latter now prints a one-line
transcript hint suggesting az login and a re-run, then falls
through to the existing manual prompts. Operators who would
have benefited from the find/create shortcut now know it is
one az login away.
trollbridge init interactive wizard now offers to find or
create an Azure OpenAI deployment via the az CLI when the
operator picks aoai as the provider (#132). Detection: az
in PATH AND az account show succeeds; otherwise the wizard
silently falls back to the manual endpoint / key prompts. The find branch lists the operator's existing OpenAI accounts and
deployments and pre-fills endpoint / model / (user-mode) API
key from the selection. The create branch walks them through
resource-group + account-name + region (default eastus) +
deployment-name and provisions via az cognitiveservices
commands. Operators who want the manual flow pick skip. Az
must be logged in to the desired subscription before init
starts (az account set --subscription <id> if multiple).
make llm-test (new) runs the LLM-advisor regression suite
against the live LLM configured by your trollbridge.yaml
(closes #133). Point TROLLBRIDGE_LLM_TEST_CONFIG at the
config, then make llm-test. Bundles under llmtest/bundles/*.yaml declare directives + allow/deny
context + cases with expected verdict (allow/deny/ ask_user) and confidence band (low/medium/high). The
framework dispatches one live LLM call per case and reports
per-case pass/fail — catches prompt drift, model-version
drift, and subtle policy gaps. Three starter bundles
(baseline / security / grey-area) ship by default; add your
own under llmtest/bundles/. See llmtest/README.md.
Forensics
Held requests resolved after signal_after_seconds fires now write
a follow-up audit entry sharing the original request_id (#97).
The entry carries post_signal_resolution: true and signal_after_seconds: <N> (new omitempty fields on audit.Entry; audit_schema_version stays at 1).
logging.audit_level knob (#113). Three levels: all (default;
current behavior), decisions (only entries from a human or the
LLM advisor — static-policy auto-decisions dropped at enqueue), none (drop every entry). Omitting the key preserves the
pre-#113 behavior; existing deployments are unaffected on
upgrade. Invalid values fail config validation at startup.
trollbridge logs review subcommand (#114). Lists audit entries
from human or LLM decisions in chronological order, with
reasoning and (for LLM entries) model / confidence / input-hash
trace. Static-policy auto-decisions filtered out. Shares the (DecisionSource).IsHumanOrLLM() categorization with the new audit_level: decisions filter (#113). --since <duration>
applies a cutoff; --config -c overrides the default config
path.
Docs
README rewritten as a first-time-reader front door (#135). The
page is now ~130 lines (was ~390), organized as: what trollbridge
is, an audience map ("agent is installing for me" /
"installing on my own machine" / "deploying to a host"),
install / run / verify, and a compact "Where to go next" doc map.
Operator-deep content moved to two new docs: docs/operator-ui.md (TUI keymap, daemon
mode, attach, console REPL, CI validation via trollbridge validate) and docs/self-describing.md
(config.trollbridge.dev/* bootstrap endpoints). Hosts /
CA / mTLS / TLS-interception specifics link to the existing PROXY-SETUP-AGENT.md. No CLI behavior
changed; the agentic-setup URL and curl-pipe installer
command are preserved verbatim.
CLIENT-SETUP-AGENT.md Step 3 no longer claims the README has
per-runtime trust-bundle snippets (Python certifi, Node NODE_EXTRA_CA_CERTS, Java cacerts) — it never did. The
reference now points at DESIGN.md §7.5, which does carry the
list. Embedded copy in internal/selfdescribe/ updated in
lock-step so the drift test stays green.
CLIENT-SETUP-AGENT.md Step 1 now documents two scopes for the
proxy env vars (#116). Shell-wide (existing export … pattern)
for the convenient case; per-process / agent-scoped (the VAR=val … <command> shell prefix and a PowerShell equivalent)
for agents that should not pollute the operator's interactive
shell. The trade-off is named in one line so the reader can pick
the right scope.
TUI
Approvals-pane header gains a ␇ reload failed badge when the
daemon's last hot-reload attempt errored (#129). Same red-bold
style as the existing pending-count indicator (#72) so the badge
is visible from across the room. Operators editing the config to
tighten policy now have a visual cue when their edit did not take
— the silent-divergence exposure the issue named (since strict
decoding, #123, made reload failures more likely on operator
typos) is now operator-observable, not log-only. Badge clears on
the next successful reload; no dismiss key by design. The
server's /v1/rules response carries last_reload_error, last_reload_at, last_reload_source (omitempty) — backwards-
compatible with existing consumers.
LLM panel now scrolls correctly when the selection moves below
the visible newest-first window (#117). The selection
(DigestSelected) was always moving on Up/Down/j/k — the
renderer just iterated strictly from the newest digest and
stopped at the body row budget, so past that point the operator
saw no change. Anchor-at-bottom scroll keeps the selected
digest visible; the modal-mode render (Enter-expanded, detail
doesn't fit inline) was already correct and is unaffected.